Search

The recent advent of large language models (LLMs), such as ChatGPT, has drawn attention to generative artificial intelligence (AI) in a number of fields. Generative AI can produce different types of data including text, images, and voice, depending on the training methods and datasets used. Additionally, recent advancements in multimodal techniques, which can simultaneously process multiple data types like text and images, have expanded the potential of using multimodal generative AI in the medical environment where various types of clinical and imaging information are used together. This review summarizes the concepts and types of LLMs, image generative AI, and multimodal AI, and it examines the status and future possibilities of generative AI in the field of radiology.

Abstract Accurate measurement of abdominal aortic aneurysm is essential for selecting suitable stent-grafts to avoid complications of endovascular aneurysm repair. However, the conventional image-based measurements are inaccurate and time-consuming. We introduce the automated workflow including semantic segmentation with active learning (AL) and measurement using an application programming interface of computer-aided design. 300 patients underwent CT scans, and semantic segmentation for aorta, thrombus, calcification, and vessels was performed in 60–300 cases with AL across five stages using UNETR, SwinUNETR, and nnU-Net consisted of 2D, 3D U-Net, 2D-3D U-Net ensemble, and cascaded 3D U-Net. 7 clinical landmarks were automatically measured for 96 patients. In AL stage 5, 3D U-Net achieved the highest dice similarity coefficient (DSC) with statistically significant differences (p

Abstract In this paper, we introduce in-depth the analysis of CNNs and ViT architectures in medical images, with the goal of providing insights into subsequent research direction. In particular, the origins of deep neural networks should be explainable for medical images, but there has been a paucity of studies on such explainability in the aspect of deep neural network architectures. Therefore, we investigate the origin of model performance, which is the clue to explaining deep neural networks, focusing on the two most relevant architectures, such as CNNs and ViT. We give four analyses, including (1) robustness in a noisy environment, (2) consistency in translation invariance property, (3) visual recognition with obstructed images, and (4) acquired features from shape or texture so that we compare origins of CNNs and ViT that cause the differences of visual recognition performance. Furthermore, the discrepancies between medical and generic images are explored regarding such analyses. We discover that medical images, unlike generic ones, exhibit class-sensitive. Finally, we propose a straightforward ensemble method based on our analyses, demonstrating that our findings can help build follow-up studies. Our analysis code will be publicly available.

Abstract Complex temporal bone anatomy complicates operations; thus, surgeons must engage in practice to mitigate risks, improving patient safety and outcomes. However, existing training methods often involve prohibitive costs and ethical problems. Therefore, we developed an educational mastoidectomy simulator, considering mechanical properties using 3D printing. The mastoidectomy simulator was modeled on computed tomography images of a patient undergoing a mastoidectomy. Infill was modeled for each anatomical part to provide a realistic drilling sensation. Bone and other anatomies appear in assorted colors to enhance the simulator’s educational utility. The mechanical properties of the simulator were evaluated by measuring the screw insertion torque for infill specimens and cadaveric temporal bones and investigating its usability with a five-point Likert-scale questionnaire completed by five otolaryngologists. The maximum insertion torque values of the sigmoid sinus, tegmen, and semicircular canal were 1.08 ± 0.62, 0.44 ± 0.42, and 1.54 ± 0.43 N mm, displaying similar-strength infill specimens of 40%, 30%, and 50%. Otolaryngologists evaluated the quality and usability at 4.25 ± 0.81 and 4.53 ± 0.62. The mastoidectomy simulator could provide realistic bone drilling feedback for educational mastoidectomy training while reinforcing skills and comprehension of anatomical structures.

Abstract Transfer learning plays a pivotal role in addressing the paucity of data, expediting training processes, and enhancing model performance. Nonetheless, the prevailing practice of transfer learning predominantly relies on pre-trained models designed for the natural image domain, which may not be well-suited for the medical image domain in grayscale. Recognizing the significance of leveraging transfer learning in medical research, we undertook the construction of class-balanced pediatric radiograph datasets collectively referred to as PedXnets, grounded in radiographic views using the pediatric radiographs collected over 24 years at Asan Medical Center. For PedXnets pre-training, approximately 70,000 X-ray images were utilized. Three different pre-training weights of PedXnet were constructed using Inception V3 for various radiation perspective classifications: Model-PedXnet-7C, Model-PedXnet-30C, and Model-PedXnet-68C. We validated the transferability and positive effects of transfer learning of PedXnets through pediatric downstream tasks including fracture classification and bone age assessment (BAA). The evaluation of transfer learning effects through classification and regression metrics showed superior performance of Model-PedXnets in quantitative assessments. Additionally, visual analyses confirmed that the Model-PedXnets were more focused on meaningful regions of interest.

Abstract Accurate lesion diagnosis through computed tomography (CT) and advances in laparoscopic or robotic surgeries have increased partial nephrectomy survival rates. However, accurately marking the kidney resection area through the laparoscope is a prevalent challenge. Therefore, we fabricated and evaluated a 4D-printed kidney surgical guide (4DP-KSG) for laparoscopic partial nephrectomies based on CT images. The kidney phantom and 4DP-KSG were designed based on CT images from a renal cell carcinoma patient. 4DP-KSG were fabricated using shape-memory polymers. 4DP-KSG was compressed to a 10 mm thickness and restored to simulate laparoscopic port passage. The Bland–Altman evaluation assessed 4DP-KSG shape and marking accuracies before compression and after restoration with three operators. The kidney phantom’s shape accuracy was 0.436 ± 0.333 mm, and the 4DP-KSG’s shape accuracy was 0.818 ± 0.564 mm before compression and 0.389 ± 0.243 mm after restoration, with no significant differences. The 4DP-KSG marking accuracy was 0.952 ± 0.682 mm before compression and 0.793 ± 0.677 mm after restoration, with no statistical differences between operators (p = 0.899 and 0.992). In conclusion, our 4DP-KSG can be used for laparoscopic partial nephrectomies, providing precise and quantitative kidney tumor marking between operators before compression and after restoration.

Abstract The aim of our study was to assess the performance of content-based image retrieval (CBIR) for similar chest computed tomography (CT) in obstructive lung disease. This retrospective study included patients with obstructive lung disease who underwent volumetric chest CT scans. The CBIR database included 600 chest CT scans from 541 patients. To assess the system performance, follow-up chest CT scans of 50 patients were evaluated as query cases, which showed the stability of the CT findings between baseline and follow-up chest CT, as confirmed by thoracic radiologists. The CBIR system retrieved the top five similar CT scans for each query case from the database by quantifying and comparing emphysema extent and size, airway wall thickness, and peripheral pulmonary vasculatures in descending order from the database. The rates of retrieval of the same pairs of query CT scans in the top 1–5 retrievals were assessed. Two expert chest radiologists evaluated the visual similarities between the query and retrieved CT scans using a five-point scale grading system. The rates of retrieving the same pairs of query CTs were 60.0% (30/50) and 68.0% (34/50) for top-three and top-five retrievals. Radiologists rated 64.8% (95% confidence interval 58.8–70.4) of the retrieved CT scans with a visual similarity score of four or five and at least one case scored five points in 74% (74/100) of all query cases. The proposed CBIR system for obstructive lung disease integrating quantitative CT measures demonstrated potential for retrieving chest CT scans with similar imaging phenotypes. Further refinement and validation in this field would be valuable.

Abstract The treatment decisions for patients with hepatocellular carcinoma are determined by a wide range of factors, and there is a significant difference between the recommendations of widely used staging systems and the actual initial treatment choices. Herein, we propose a machine learning-based clinical decision support system suitable for use in multi-center settings. We collected data from nine institutions in South Korea for training and validation datasets. The internal and external datasets included 935 and 1750 patients, respectively. We developed a model with 20 clinical variables consisting of two stages: the first stage which recommends initial treatment using an ensemble voting machine, and the second stage, which predicts post-treatment survival using a random survival forest algorithm. We derived the first and second treatment options from the results with the highest and the second-highest probabilities given by the ensemble model and predicted their post-treatment survival. When only the first treatment option was accepted, the mean accuracy of treatment recommendation in the internal and external datasets was 67.27% and 55.34%, respectively. The accuracy increased to 87.27% and 86.06%, respectively, when the second option was included as the correct answer. Harrell’s C index, integrated time-dependent AUC curve, and integrated Brier score of survival prediction in the internal and external datasets were 0.8381 and 0.7767, 91.89 and 86.48, 0.12, and 0.14, respectively. The proposed system can assist physicians by providing data-driven predictions for reference from other larger institutions or other physicians within the same institution when making treatment decisions.

Abstract Conventional suture anchors (CAs) and all-suture anchors (ASAs) are used for rotator cuff repair. Pull-out strength (POS) is an important factor that affects surgical outcomes. While the fixation mechanism differs between the anchor types and relies on the quality, few studies have compared biomechanical properties of anchors based on bone quality. This study aimed to compare the biomechanical properties of anchors using osteoporotic bone (OB) and non-osteoporotic bone (NOB) simulators. Humerus simulators were fabricated using fused deposition modeling of 3D printing and acrylonitrile butadiene styrene adjusting the thickness of cortical bone and density of cancellous bone based on CT images. Cyclic loading from 10 to 50 N, 10 to 100 N, and 10 to 150 N for 10 cycles was clinically determined at each anchor because the supraspinatus generates a force of 67–125 N in daily activities of normal control. After cyclic loading, the anchor was extruded at a load of 5 mm/min. Displacement, POS, and stiffness were measured. In OB simulators, CAs revealed bigger gap displacement than ASAs with cyclic loading of 10–150 N. ASA showed higher values for POS and stiffness. In NOB simulators, ASAs revealed bigger gap displacement than CAs with cyclic loading of 10–150 N. ASA showed higher values for POS and CA showed higher values for stiffness. POS of anchors depends on anchors ‘displacement and bone stiffness. In conclusion, ASA demonstrated better biomechanical performance than CA in terms of stability under cyclic loading and stiffness with similar POS in OB.

Abstract The lateral cephalogram in orthodontics is a valuable screening tool on undetected obstructive sleep apnea (OSA), which can lead to consequences of severe systematic disease. We hypothesized that a deep learning-based classifier might be able to differentiate OSA as anatomical features in lateral cephalogram. Moreover, since the imaging devices used by each hospital could be different, there is a need to overcome modality difference of radiography. Therefore, we proposed a deep learning model with knowledge distillation to classify patients into OSA and non-OSA groups using the lateral cephalogram and to overcome modality differences simultaneously. Lateral cephalograms of 500 OSA patients and 498 non-OSA patients from two different devices were included. ResNet-50 and ResNet-50 with a feature-based knowledge distillation models were trained and their performances of classification were compared. Through the knowledge distillation, area under receiver operating characteristic curve analysis and gradient-weighted class activation mapping of knowledge distillation model exhibits high performance without being deceived by features caused by modality differences. By checking the probability values predicting OSA, an improvement in overcoming the modality differences was observed, which could be applied in the actual clinical situation.

Abstract The study aimed to identify critical factors associated with the surgical stability of pogonion (Pog) by applying machine learning (ML) to predict relapse following two-jaw orthognathic surgery (2 J-OGJ). The sample set comprised 227 patients (110 males and 117 females, 207 training and 20 test sets). Using lateral cephalograms taken at the initial evaluation (T0), pretreatment (T1), after (T2) 2 J-OGS, and post treatment (T3), 55 linear and angular skeletal and dental surgical movements (T2-T1) were measured. Six ML modes were utilized, including classification and regression trees (CART), conditional inference tree (CTREE), and random forest (RF). The training samples were classified into three groups; highly significant (HS) (≥ 4), significant (S) (≥ 2 and

Abstract Partial nephrectomy has been demonstrated to preserve renal function compared with radical nephrectomy. Computed tomography (CT) is used to reveal localized renal cell carcinoma (RCC). However, marking RCC directly and quantitatively on a patient's kidney during an operation is difficult. We fabricated and evaluated a 3D-printed kidney surgical guide (3DP-KSG) with a realistic kidney phantom. The kidney phantoms including parenchyma and three different RCC locations and 3DP-KSG were designed and fabricated based on a patient's CT image. 3DP-KSG was used to insert 16-gauge intravenous catheters into the kidney phantoms, which was scanned by CT. The catheter insertion points and angle were evaluated. The measurement errors of insertion points were 1.597 ± 0.741 mm, and cosine similarity of trajectories was 0.990 ± 0.010. The measurement errors for X-axis, Y-axis, and Z-axis in the insertion point were 0.611 ± 0.855 mm, 0.028 ± 1.001 mm, and − 0.510 ± 0.923 mm. The 3DP-KSG targeted the RCC accurately, quantitatively, and immediately on the surface of the kidney, and no significant difference was shown between the operators. Partial nephrectomy will accurately remove the RCC using 3DP-KSG in the operating room.

Abstract Accurate and reliable detection of intracranial aneurysms is vital for subsequent treatment to prevent bleeding. However, the detection of intracranial aneurysms can be time-consuming and even challenging, and there is great variability among experts, especially in the case of small aneurysms. This study aimed to detect intracranial aneurysms accurately using a convolutional neural network (CNN) with 3D time-of-flight magnetic resonance angiography (TOF-MRA). A total of 154 3D TOF-MRA datasets with intracranial aneurysms were acquired, and the gold standards were manually drawn by neuroradiologists. We also obtained 113 subjects from a public dataset for external validation. These angiograms were pre-processed by using skull-stripping, signal intensity normalization, and N4 bias correction. The 3D patches along the vessel skeleton from MRA were extracted. Values of the ratio between the aneurysmal and the normal patches ranged from 1:1 to 1:5. The semantic segmentation on intracranial aneurysms was trained using a 3D U-Net with an auxiliary classifier to overcome the imbalance in patches. The proposed method achieved an accuracy of 0.910 in internal validation and external validation accuracy of 0.883 with a 2:1 ratio of normal to aneurysmal patches. This multi-task learning method showed that the aneurysm segmentation performance was sufficient to be helpful in an actual clinical setting.

Abstract Semantic segmentation of breast and surrounding tissues in supine and prone breast magnetic resonance imaging (MRI) is required for various kinds of computer-assisted diagnoses for surgical applications. Variability of breast shape in supine and prone poses along with various MRI artifacts makes it difficult to determine robust breast and surrounding tissue segmentation. Therefore, we evaluated semantic segmentation with transfer learning of convolutional neural networks to create robust breast segmentation in supine breast MRI without considering supine or prone positions. Total 29 patients with T1-weighted contrast-enhanced images were collected at Asan Medical Center and two types of breast MRI were performed in the prone position and the supine position. The four classes, including lungs and heart, muscles and bones, parenchyma with cancer, and skin and fat, were manually drawn by an expert. Semantic segmentation on breast MRI scans with supine, prone, transferred from prone to supine, and pooled supine and prone MRI were trained and compared using 2D U-Net, 3D U-Net, 2D nnU-Net and 3D nnU-Net. The best performance was 2D models with transfer learning. Our results showed excellent performance and could be used for clinical purposes such as breast registration and computer-aided diagnosis.

Abstract The validation of the accuracy of the quantification software in computed tomography (CT) images is very challenging. Therefore, we proposed a CT imaging phantom that accurately represents patient-specific anatomical structures and randomly integrates various lesions including disease-like patterns and lesions of various shapes and sizes using silicone casting and three-dimensional (3D) printing. Six nodules of various shapes and sizes were randomly added to the patient’s modeled lungs to evaluate the accuracy of the quantification software. By using silicone materials, CT intensities suitable for the lesions and lung parenchyma were realized, and their Hounsfield unit (HU) values were evaluated on a CT scan of the phantom. As a result, based on the CT scan of the imaging phantom model, the measured HU values for the normal lung parenchyma, each nodule, fibrosis, and emphysematous lesions were within the target value. The measurement error between the stereolithography model and 3D-printing phantoms was 0.2 ± 0.18 mm. In conclusion, the use of 3D printing and silicone casting allowed the application and evaluation of the proposed CT imaging phantom for the validation of the accuracy of the quantification software in CT images, which could be applied to CT-based quantification and development of imaging biomarkers.

Abstract Breast cancer is a common cancer among women, and screening mammography is the primary tool for diagnosing this condition. Recent advancements in deep-learning technologies have triggered the implementation of research studies via mammography. Semi-supervised or unsupervised methods are often used to overcome the limitations of supervised learning, such as manpower and time, for labeling in clinical situations where abnormal data are significantly lacking. Accordingly, we proposed a generative model that uses a state-of-the-art generative network (StyleGAN2) to create high-quality synthetic mammographic images and an anomaly detection method to detect breast cancer on mammograms in unsupervised methods. The generation model was trained via only normal mammograms and breast cancer classification was performed via anomaly detection using 50 breast cancer and 50 normal mammograms that did not overlap with the dataset for generative model learning. Our generative model has shown comparable fidelity to real images, and the anomaly detection method via this generative model showed high sensitivity, demonstrating its potential for breast cancer screening. This method could differentiate between normal and cancer-positive mammogram and help overcome the weakness of current supervised methods.

Abstract The generative adversarial network (GAN) is a promising deep learning method for generating images. We evaluated the generation of highly realistic and high-resolution chest radiographs (CXRs) using progressive growing GAN (PGGAN). We trained two PGGAN models using normal and abnormal CXRs, solely relying on normal CXRs to demonstrate the quality of synthetic CXRs that were 1000 × 1000 pixels in size. Image Turing tests were evaluated by six radiologists in a binary fashion using two independent validation sets to judge the authenticity of each CXR, with a mean accuracy of 67.42% and 69.92% for the first and second trials, respectively. Inter-reader agreements were poor for the first (κ = 0.10) and second (κ = 0.14) Turing tests. Additionally, a convolutional neural network (CNN) was used to classify normal or abnormal CXR using only real images and/or synthetic images mixed datasets. The accuracy of the CNN model trained using a mixed dataset of synthetic and real data was 93.3%, compared to 91.0% for the model built using only the real data. PGGAN was able to generate CXRs that were identical to real CXRs, and this showed promise to overcome imbalances between classes in CNN training.

Recent studies of automatic diagnosis of vertebral compression fractures (VCFs) using deep learning mainly focus on segmentation and vertebral level detection in lumbar spine lateral radiographs (LSLRs). Herein, we developed a model for simultaneous VCF diagnosis and vertebral level detection without using adjacent vertebral bodies. In total, 1102 patients with VCF, 1171 controls were enrolled. The 1865, 208, and 198 LSLRS were divided into training, validation, and test dataset. A ground truth label with a 4-point trapezoidal shape was made based on radiological reports showing normal or VCF at some vertebral level. We applied a modified U-Net architecture, in which decoders were trained to detect VCF and vertebral levels, sharing the same encoder. The multi-task model was significantly better than the single-task model in sensitivity and area under the receiver operating characteristic curve. In the internal dataset, the accuracy, sensitivity, and specificity of fracture detection per patient or vertebral body were 0.929, 0.944, and 0.917 or 0.947, 0.628, and 0.977, respectively. In external validation, those of fracture detection per patient or vertebral body were 0.713, 0.979, and 0.447 or 0.828, 0.936, and 0.820, respectively. The success rates were 96 % and 94 % for vertebral level detection in internal and external validation, respectively. The multi-task-shared encoder was significantly better than the single-task encoder. Furthermore, both fracture and vertebral level detection was good in internal and external validation. Our deep learning model may help radiologists perform real-life medical examinations.

Abstract Recently, the development of 3D printing (3DP) technology and its application in various fields have improved our quality of life. However, hazardous materials that affect the human body, such as formaldehyde and particulate matter (PM), are emitted into the air during 3DP. This study measured the formaldehyde, PM10, and PM2.5 emitted by 3DP with the ventilation operation using six materials in material extrusion (ME) and vat photopolymerization (VP) and compared them between the 3DP workspace and the control setting with test–retest validation by two researchers. The experiments were divided into four stages based on the 3DP and ventilation operation. A linear mixed model was used to analyze the mean differences and tendencies between the 3DP workspace and the control setting. The change as ventilation was switched from off to on was evaluated by calculating the area. The differences and tendencies were shown in the statistically significant differences from a post-hoc test (α = 0.0125) except for some cases. There was a significant difference in formaldehyde depending on the ventilation operation; however, only a minor difference in PM10, and PM2.5 was confirmed. The amount of formaldehyde exceeding the standard was measured in all materials during 3DP without ventilation. Therefore, it is recommended to operate ventilation systems.

Abstract The aim of this study was to develop an auto-segmentation algorithm for mandibular condyle using the 3D U-Net and perform a stress test to determine the optimal dataset size for achieving clinically acceptable accuracy. 234 cone-beam computed tomography images of mandibular condyles were acquired from 117 subjects from two institutions, which were manually segmented to generate the ground truth. Semantic segmentation was performed using basic 3D U-Net and a cascaded 3D U-Net. A stress test was performed using different sets of condylar images as the training, validation, and test datasets. Relative accuracy was evaluated using dice similarity coefficients (DSCs) and Hausdorff distance (HD). In the five stages, the DSC ranged 0.886–0.922 and 0.912–0.932 for basic 3D U-Net and cascaded 3D U-Net, respectively; the HD ranged 2.557–3.099 and 2.452–2.600 for basic 3D U-Net and cascaded 3D U-Net, respectively. Stage V (largest data from two institutions) exhibited the highest DSC of 0.922 ± 0.021 and 0.932 ± 0.023 for basic 3D U-Net and cascaded 3D U-Net, respectively. Stage IV (200 samples from two institutions) had a lower performance than stage III (162 samples from one institution). Our results show that fully automated segmentation of mandibular condyles is possible using 3D U-Net algorithms, and the segmentation accuracy increases as training data increases.

Purpose To develop and validate a deep learning-based screening tool for the early diagnosis of scoliosis using chest radiographs with a semi-supervised generative adversarial network (GAN). Materials and Methods Using a semi-supervised learning framework with a GAN, a screening tool for diagnosing scoliosis was developed and validated through the chest PA radiographs of patients at two different tertiary hospitals. Our proposed method used training GAN with mild to severe scoliosis only in a semi-supervised manner, as an upstream task to learn scoliosis representations and a downstream task to perform simple classification for differentiating between normal and scoliosis states sensitively. Results The area under the receiver operating characteristic curve, negative predictive value (NPV), positive predictive value, sensitivity, and specificity were 0.856, 0.950, 0.579, 0.985, and 0.285, respectively. Conclusion Our deep learning-based artificial intelligence software in a semi-supervised manner achieved excellent performance in diagnosing scoliosis using the chest PA radiographs of young individuals; thus, it could be used as a screening tool with high NPV and sensitivity and reduce the burden on radiologists for diagnosing scoliosis through health screening chest radiographs.

Abstract The detection of Epstein–Barr virus (EBV) in gastric cancer patients is crucial for clinical decision making, as it is related with specific treatment responses and prognoses. Despite its importance, the limited medical resources preclude universal EBV testing. Herein, we propose a deep learning-based EBV prediction method from H&E-stained whole-slide images (WSI). Our model was developed using 319 H&E stained WSI (26 EBV positive; TCGA dataset) from the Cancer Genome Atlas, and 108 WSI (8 EBV positive; ISH dataset) from an independent institution. Our deep learning model, EBVNet consists of two sequential components: a tumor classifier and an EBV classifier. We visualized the learned representation by the classifiers using UMAP. We externally validated the model using 60 additional WSI (7 being EBV positive; HGH dataset). We compared the model’s performance with those of four pathologists. EBVNet achieved an AUPRC of 0.65, whereas the four pathologists yielded a mean AUPRC of 0.41. Moreover, EBVNet achieved an negative predictive value, sensitivity, specificity, precision, and F1-score of 0.98, 0.86, 0.92, 0.60, and 0.71, respectively. Our proposed model is expected to contribute to prescreen patients for confirmatory testing, potentially to save test-related cost and labor.

Abstract Realistic image synthesis based on deep learning is an invaluable technique for developing high-performance computer aided diagnosis systems while protecting patient privacy. However, training a generative adversarial network (GAN) for image synthesis remains challenging because of the large amounts of data required for training various kinds of image features. This study aims to synthesize retinal images indistinguishable from real images and evaluate the efficacy of the synthesized images having a specific disease for augmenting class imbalanced datasets. The synthesized images were validated via image Turing tests, qualitative analysis by retinal specialists, and quantitative analyses on amounts and signal-to-noise ratios of vessels. The efficacy of synthesized images was verified by deep learning-based classification performance. Turing test shows that accuracy, sensitivity, and specificity of 54.0 ± 12.3%, 71.1 ± 18.8%, and 36.9 ± 25.5%, respectively. Here, sensitivity represents correctness to find real images among real datasets. Vessel amounts and average SNR comparisons show 0.43% and 1.5% difference between real and synthesized images. The classification performance after augmenting synthesized images outperforms every ratio of imbalanced real datasets. Our study shows the realistic retina images were successfully generated with insignificant differences between the real and synthesized images and shows great potential for practical applications.

Triage is essential for the early diagnosis and reporting of emergency patients in the emergency department. Here, the authors develop an anomaly detection algorithm with a deep generative model that reprioritizes radiology worklists and provides lesion attention maps for brain CT images with critical findings.

BackgroundSurgical reconstruction of the urinary tract, anus, and vagina is the definitive treatment for cloacal malformation. However, this procedure may be technically challenging in patients with a long common channel (>3 cm), because further reconstructive procedures, such as vaginal replacement or vaginal switch maneuver, may be required. Thus, accurate determination of spatial anatomy is essential during surgical planning. Three-dimensional (3D) reconstruction using rotational fluoroscopy, computed tomography (CT), and magnetic resonance imaging (MRI) has recently been reported to help in determining the relationship between the rectum, vagina, and bladder, and provides a more accurate measurement of the channel length compared to conventional cloacography. MRI-based 3D reconstruction provides substantial information regarding soft tissue structures around the cloaca, including the pelvic floor musculature and anus.CaseA 2-year-old girl with cloacal malformation required reconstructive surgery. Colostomy and cystostomy had been performed on the first day of her life. Preoperative loopogram revealed a cloaca with a long common channel (35 mm) and short urethra (9 mm), single vaginal opening in the bladder neck, and the colon anterior to the vagina with a fistula at the vaginal neck. Because the vagina was too short to be pulled through, 3D printing based on MRI was performed to visualize structural relationships prior to surgical correction. Saline was used for cloacal visualization. Furthermore, endoscopy-assisted urogenital mobilization was performed, and vaginal substitution was performed using the rectum. No postoperative complications were observed.ConclusionsWe believe this is the first report of the use of MRI-based 3D imaging and printing, with saline as a contrast agent during surgical planning for correction of cloacal malformation. MRI-based 3D printing is a potentially promising technique for surgical planning of cloacal malformation correction in patients with a long common channel, as it provides detailed information about the surrounding soft tissue structures without exposure to radiation or contrasting agents.

Abstract Left atrial appendage (LAA) occlusion (LAAO) is used to close the finger-like extension from the left atrium with occlusion devices to block the source of thrombosis. However, selection of the devices size is not easy due to various anatomical changes. The purpose of this study is patient-specific, computed tomography angiography (CTA)-based, three-dimensionally (3D) printed LAAO phantoms were applied pre-procedure to determine the size. Ten patients were enrolled prospectively in March 2019 and December 2020. The cardiac structure appearing in CTA was first segmented, and the left atrium and related structures in the LAAO procedure were modeled. The phantoms were fabricated using two methods of fused deposition modeling (FDM) and stereolithography (SLA) 3D printers with thermoplastic polyurethane (TPU) and flexible resin materials and evaluated by comparing their physical and material properties. The 3D-printed phantoms were directly used to confirm the shape of LAA, and to predict the device size for LAAO. In summary, the shore A hardness of TPU of FDM was about 80–85 shore A, and that of flexible resin of SLA was about 50–70 shore A. The measurement error between the STL model and 3D printing phantoms were 0.45 ± 0.37 mm (Bland–Altman, limits of agreement from − 1.8 to 1.6 mm). At the rehearsal, the estimations of device sizes were the exact same with those in the actual procedures of all 10 patients. In conclusion, simulation with a 3D-printed left atrium phantom could be used to predict the LAAO insertion device size accurately before the procedure.

ObjectiveConventional computer-aided diagnosis using convolutional neural networks (CNN) has limitations in detecting sensitive changes and determining accurate decision boundaries in spectral and structural diseases such as scoliosis. We devised a new method to detect and diagnose adolescent idiopathic scoliosis in chest X-rays (CXRs) employing the latent space's discriminative ability in the generative adversarial network (GAN) and a simple multi-layer perceptron (MLP) to screen adolescent idiopathic scoliosis CXRs.Materials and methodsOur model was trained and validated in a two-step manner. First, we trained a GAN using CXRs with various scoliosis severities and utilized the trained network as a feature extractor using the GAN inversion method. Second, we classified each vector from the latent space using a simple MLP.ResultsThe 2-layer MLP exhibited the best classification in the ablation study. With this model, the area under the receiver operating characteristic (AUROC) curves were 0.850 in the internal and 0.847 in the external datasets. Furthermore, when the sensitivity was fixed at 0.9, the model's specificity was 0.697 in the internal and 0.646 in the external datasets.ConclusionWe developed a classifier for Adolescent idiopathic scoliosis (AIS) through generative representation learning. Our model shows good AUROC under screening chest radiographs in both the internal and external datasets. Our model has learned the spectral severity of AIS, enabling it to generate normal images even when trained solely on scoliosis radiographs.

Abstract Despite being the gold standard for diagnosis of osteoporosis, dual-energy X-ray absorptiometry (DXA) could not be widely used as a screening tool for osteoporosis. This study aimed to predict osteoporosis via simple hip radiography using deep learning algorithm. A total of 1001 datasets of proximal femur DXA with matched same-side cropped simple hip bone radiographic images of female patients aged ≥ 55 years were collected. Of these, 504 patients had osteoporosis (T-score ≤ − 2.5), and 497 patients did not have osteoporosis. The 1001 images were randomly divided into three sets: 800 images for the training, 100 images for the validation, and 101 images for the test. Based on VGG16 equipped with nonlocal neural network, we developed a deep neural network (DNN) model. We calculated the confusion matrix and evaluated the accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). We drew the receiver operating characteristic (ROC) curve. A gradient-based class activation map (Grad-CAM) overlapping the original image was also used to visualize the model performance. Additionally, we performed external validation using 117 datasets. Our final DNN model showed an overall accuracy of 81.2%, sensitivity of 91.1%, and specificity of 68.9%. The PPV was 78.5%, and the NPV was 86.1%. The area under the ROC curve value was 0.867, indicating a reasonable performance for screening osteoporosis by simple hip radiography. The external validation set confirmed a model performance with an overall accuracy of 71.8% and an AUC value of 0.700. All Grad-CAM results from both internal and external validation sets appropriately matched the proximal femur cortex and trabecular patterns of the radiographs. The DNN model could be considered as one of the useful screening tools for easy prediction of osteoporosis in the real-world clinical setting.

Abstract In thoracoabdominal aortic aneurysm repair, repairing the visceral and segmental arteries is challenging. Although there is a pre-hand-sewn and multi-branched graft based on the conventional image-based technique, it has shortcomings in precisely positioning and directing the visceral and segmental arteries. Here, we introduce two new reconstruction techniques using patient-specific 3D-printed graft reconstruction guides: (1) model-based technique that presents the projected aortic graft, visualizing the main aortic body and its major branches and (2) guide-based technique in which the branching vessels in the visualization model are replaced by marking points identifiable by tactile sense. We demonstrate the effectiveness by evaluating conventional and new techniques based on accuracy, marking time requirement, reproducibility, and results of survey to surgeons on the perceived efficiency and efficacy. The graft reconstruction guides cover the segmentation, design, fabrication, post-processing, and clinical application of open surgical repair of thoracoabdominal aneurysm, and proved to be efficient for accurately reconstructing customized grafts.

Abstract Heterogeneous clinical manifestations and progression of chronic obstructive pulmonary disease (COPD) affect patient health risk assessment, stratification, and management. Pulmonary function tests are used to diagnose and classify the severity of COPD, but they cannot fully represent the type or range of pathophysiologic abnormalities of the disease. To evaluate whether deep radiomics from chest computed tomography (CT) images can predict mortality in patients with COPD, we designed a convolutional neural network (CNN) model for extracting representative features from CT images and then performed random survival forest to predict survival in COPD patients. We trained CNN-based binary classifier based on six-minute walk distance results (> 440 m or not) and extracted high-throughput image features (i.e., deep radiomics) directly from the last fully connected layer of it. The various sizes of fully connected layers and combinations of deep features were experimented using a discovery cohort with 344 patients from the Korean Obstructive Lung Disease cohort and an external validation cohort with 102 patients from Penang General Hospital in Malaysia. In the integrative analysis of discovery and external validation cohorts, with combining 256 deep features from the coronal slice of the vertebral body and two sagittal slices of the left/right lung, deep radiomics for survival prediction achieved concordance indices of 0.8008 (95% CI, 0.7642–0.8373) and 0.7156 (95% CI, 0.7024–0.7288), respectively. Deep radiomics from CT images could be used to predict mortality in COPD patients.

Abstract Realistic image generation is valuable in dental medicine, but still challenging for generative adversarial networks (GANs), which require large amounts of data to overcome the training instability. Thus, we generated lateral cephalogram X-ray images using a deep-learning-based progressive growing GAN (PGGAN). The quality of generated images was evaluated by three methods. First, signal-to-noise ratios of real/synthesized images, evaluated at the posterior arch region of the first cervical vertebra, showed no statistically significant difference (t-test, p = 0.211). Second, the results of an image Turing test, conducted by non-orthodontists and orthodontists for 100 randomly chosen images, indicated that they had difficulty in distinguishing whether the image was real or synthesized. Third, cephalometric tracing with 42 landmark points detection, performed on real and synthesized images by two expert orthodontists, showed consistency with mean difference of 2.08 ± 1.02 mm. Furthermore, convolutional neural network-based classification tasks were used to classify skeletal patterns using a real dataset with class imbalance and a dataset balanced with synthesized images. The classification accuracy for the latter case was increased by 1.5%/3.3% at internal/external test sets, respectively. Thus, the cephalometric images generated by PGGAN are sufficiently realistic and have potential to application in various fields of dental medicine.

Abstract Generative adversarial network (GAN) creates synthetic images to increase data quantity, but whether GAN ensures meaningful morphologic variations is still unknown. We investigated whether GAN-based synthetic images provide sufficient morphologic variations to improve molecular-based prediction, as a rare disease of isocitrate dehydrogenase (IDH)-mutant glioblastomas. GAN was initially trained on 500 normal brains and 110 IDH-mutant high-grade astocytomas, and paired contrast-enhanced T1-weighted and FLAIR MRI data were generated. Diagnostic models were developed from real IDH-wild type (n = 80) with real IDH-mutant glioblastomas (n = 38), or with synthetic IDH-mutant glioblastomas, or augmented by adding both real and synthetic IDH-mutant glioblastomas. Turing tests showed synthetic data showed reality (classification rate of 55%). Both the real and synthetic data showed that a more frontal or insular location (odds ratio [OR] 1.34 vs. 1.52; P = 0.04) and distinct non-enhancing tumor margins (OR 2.68 vs. 3.88; P

Abstract In children with mandibular hypoplasia, airway management is challenging. However, detailed cephalometric assessment data for this population are sparse. The aim of this study was to find risk factors for predicting difficult airways in children with mandibular hypoplasia, and compare upper airway anatomical differences using three-dimensional computed tomography (3D CT) between children with mandibular hypoplasia and demographically matched healthy controls. There were significant discrepancies in relative tongue position (P

Abstract Skin cancer usually occurs in the facial area relatively exposed to sunlight. Medical imaging can confirm the invasiveness and metastasis of skin cancer, which is used to establish a surgical plan. However, there is no method of directly marking this information on the patient's skin in the operating room. We evaluated a skin cancer resection guide that marks resection areas including safety margins on the patient's skin based on medical images and in-vitro phantom fabricated via 3D printing. The in-vitro phantom, which includes the skull, skin, and five different cancer locations was designed and fabricated based on a CT image of a patient. Skin cancer resection guides were designed using a CT image of an in-vitro phantom, with a safety margin, and four injection points at each cancer. The guide was used to insert 16 cc intravenous catheters into each cancer of the phantom, which was rescanned by CT. The catheter insertion point and angle were evaluated. The accuracy of the insertion points was 2.09 ± 1.06 mm and cosine similarities was 0.980 ± 0.020. In conclusion, skin cancer resection guides were fabricated to mark surgical plans on the patient's skin in the operating room. They demonstrated reasonable accuracies in actual clinical settings.

Abstract The quality of cephalometric analysis depends on the accuracy of the delineating landmarks in orthodontic and maxillofacial surgery. Due to the extensive number of landmarks, each analysis costs orthodontists considerable time per patient, leading to fatigue and inter- and intra-observer variabilities. Therefore, we proposed a fully automated cephalometry analysis with a cascade convolutional neural net (CNN). One thousand cephalometric x-ray images (2 k × 3 k) pixel were used. The dataset was split into training, validation, and test sets as 8:1:1. The 43 landmarks from each image were identified by an expert orthodontist. To evaluate intra-observer variabilities, 28 images from the dataset were randomly selected and measured again by the same orthodontist. To improve accuracy, a cascade CNN consisting of two steps was used for transfer learning. In the first step, the regions of interest (ROIs) were predicted by RetinaNet. In the second step, U-Net detected the precise landmarks in the ROIs. The average error of ROI detection alone was 1.55 ± 2.17 mm. The model with the cascade CNN showed an average error of 0.79 ± 0.91 mm (paired t-test, p = 0.0015). The orthodontist’s average error of reproducibility was 0.80 ± 0.79 mm. An accurate and fully automated cephalometric analysis was successfully developed and evaluated.

Abstract Herein, realistic and reusable phantoms for simulation of pediatric lung video-assisted thoracoscopic surgery (VATS) were proposed and evaluated. 3D-printed phantoms for VATS were designed based on chest computed tomography (CT) data of a pediatric patient with esophageal atresia and tracheoesophageal fistula. Models reflecting the patient-specific structure were fabricated based on the CT images. Appropriate reusable design, realistic mechanical properties with various material types, and 3D printers (fused deposition modeling (FDM) and PolyJet printers) were used to represent the realistic anatomical structures. As a result, the phantom printed by PolyJet reflected closer mechanical properties than those of the FDM phantom. Accuracies (mean difference ± 95 confidence interval) of phantoms by FDM and PolyJet were 0.53 ± 0.46 and 0.98 ± 0.55 mm, respectively. Phantoms were used by surgeons for VATS training, which is considered more reflective of the clinical situation than the conventional simulation phantom. In conclusion, the patient-specific, realistic, and reusable VATS phantom provides a better understanding the complex anatomical structure of a patient and could be used as an educational phantom for esophageal structure replacement in VATS.

Abstract We used 3D printed-breast surgical guides (3DP-BSG) to designate the original tumor area from the pre-treatment magnetic resonance imaging (MRI) during breast-conserving surgery (BCS) in breast cancer patients who received neoadjuvant systemic therapy (NST). Targeting the original tumor area in such patients using conventional localization techniques is difficult. For precise BCS, a method that marks the tumor area found on MRI directly to the breast is needed. In this prospective study, patients were enrolled for BCS after receiving NST. Partial resection was performed using a prone/supine MRI-based 3DP-BSG. Frozen biopsies were analyzed to confirm clear tumor margins. The tumor characteristics, pathologic results, resection margins, and the distance between the tumor and margin were analyzed. Thirty-nine patients were enrolled with 3DP-BSG for BCS. The median nearest distance between the tumor and the resection margin was 3.9 cm (range 1.2–7.8 cm). Frozen sections showed positive margins in 4/39 (10.3%) patients. Three had invasive cancers, and one had carcinoma in situ; all underwent additional resection. Final pathology revealed clear margins. After 3-year surveillance, 3/39 patients had recurrent breast cancer. With 3DP-BSG for BCS in breast cancer patients receiving NST, the original tumor area can be identified and marked directly on the breast, which is useful for surgery. Trial Registration: Clinical Research Information Service (CRIS) Identifier Number: KCT0002272. First registration number and date: No. 1 (27/04/2016).

Abstract The endoscopic features between herpes simplex virus (HSV) and cytomegalovirus (CMV) esophagitis overlap significantly, and hence the differential diagnosis between HSV and CMV esophagitis is sometimes difficult. Therefore, we developed a machine-learning-based classifier to discriminate between CMV and HSV esophagitis. We analyzed 87 patients with HSV esophagitis and 63 patients with CMV esophagitis and developed a machine-learning-based artificial intelligence (AI) system using a total of 666 endoscopic images with HSV esophagitis and 416 endoscopic images with CMV esophagitis. In the five repeated five-fold cross-validations based on the hue–saturation–brightness color model, logistic regression with a least absolute shrinkage and selection operation showed the best performance (sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and area under the receiver operating characteristic curve: 100%, 100%, 100%, 100%, 100%, and 1.0, respectively). Previous history of transplantation was included in classifiers as a clinical factor; the lower the performance of these classifiers, the greater the effect of including this clinical factor. Our machine-learning-based AI system for differential diagnosis between HSV and CMV esophagitis showed high accuracy, which could help clinicians with diagnoses.

Abstract To facilitate precise tumor resection at the time of breast-conserving surgery (BCS), we developed and implemented a magnetic resonance imaging (MRI)-based three-dimensional-printed (3DP) breast surgical guide (BSG). This prospective cohort study was conducted at a single institution from July 2017 to February 2019 on women with breast cancer who underwent partial breast resection using patient-specific 3DP BSGs. Eighty-eight patients with invasive cancer were enrolled, of whom 1 patient had bilateral breast cancer. The mean size of the tumor long-axis on MRI before surgery was 2.8 ± 0.9 cm, and multiple tumors were observed in 34 patients. In 16 cases (18.0%), the resection margin was tumor-positive according to intraoperative frozen biopsy; all of these tumors were ductal carcinoma in situ and were re-excised intraoperatively. In 93.3% of the cases, the resection margin was tumor-free in the permanent pathology. The mean pathological tumor size was 1.7 ± 1.0 cm, and the mean distance from the tumor to the border was 1.5 ± 1.0 cm. This exploratory study showed that the tumor area on the MRI could be directly displayed on the breast when using a 3DP BSG for BCS, thereby allowing precise surgery and safe tumor removal. Trial Registration Clinical Research Information Service (CRIS) Identifier (No. KCT0002375, KCT0003043).

Abstract Fast and accurate confirmation of metastasis on the frozen tissue section of intraoperative sentinel lymph node biopsy is an essential tool for critical surgical decisions. However, accurate diagnosis by pathologists is difficult within the time limitations. Training a robust and accurate deep learning model is also difficult owing to the limited number of frozen datasets with high quality labels. To overcome these issues, we validated the effectiveness of transfer learning from CAMELYON16 to improve performance of the convolutional neural network (CNN)-based classification model on our frozen dataset (N = 297) from Asan Medical Center (AMC). Among the 297 whole slide images (WSIs), 157 and 40 WSIs were used to train deep learning models with different dataset ratios at 2, 4, 8, 20, 40, and 100%. The remaining, i.e., 100 WSIs, were used to validate model performance in terms of patch- and slide-level classification. An additional 228 WSIs from Seoul National University Bundang Hospital (SNUBH) were used as an external validation. Three initial weights, i.e., scratch-based (random initialization), ImageNet-based, and CAMELYON16-based models were used to validate their effectiveness in external validation. In the patch-level classification results on the AMC dataset, CAMELYON16-based models trained with a small dataset (up to 40%, i.e., 62 WSIs) showed a significantly higher area under the curve (AUC) of 0.929 than those of the scratch- and ImageNet-based models at 0.897 and 0.919, respectively, while CAMELYON16-based and ImageNet-based models trained with 100% of the training dataset showed comparable AUCs at 0.944 and 0.943, respectively. For the external validation, CAMELYON16-based models showed higher AUCs than those of the scratch- and ImageNet-based models. Model performance for slide feasibility of the transfer learning to enhance model performance was validated in the case of frozen section datasets with limited numbers.

Abstract Background The association between long-term exposure to air pollutants, including nitrogen dioxide (NO2), carbon monoxide (CO), sulfur dioxide (SO2), ozone (O3), and particulate matter 10 μm or less in diameter (PM10), and mortality by ischemic heart disease (IHD), cerebrovascular disease (CVD), pneumonia (PN), and chronic lower respiratory disease (CLRD) is unclear. We investigated whether living in an administrative district with heavy air pollution is associated with an increased risk of mortality by the diseases through an ecological study using South Korean administrative data over 19 years. Methods A total of 249 Si-Gun-Gus, unit of administrative districts in South Korea were studied. In each district, the daily concentrations of CO, SO2, NO2, O3, and PM10 were averaged over 19 years (2001–2018). Age-adjusted mortality rates by IHD, CVD, PN and CLRD for each district were averaged for the same study period. Multivariate beta-regression analysis was performed to estimate the associations between air pollutant concentrations and mortality rates, after adjusting for confounding factors including altitude, population density, higher education rate, smoking rate, obesity rate, and gross regional domestic product per capita. Associations were also estimated for two subgrouping schema: Capital and non-Capital areas (77:172 districts) and urban and rural areas (168:81 districts). Results For IHD, higher SO2 concentrations were significantly associated with a higher mortality rate, whereas other air pollutants had null associations. For CVD, SO2 and PM10 concentrations were significantly associated with a higher mortality rate. For PN, O3 concentrations had significant positive associations with a higher mortality rate, while SO2, NO2, and PM10 concentrations had significant negative associations. For CLRD, O3 concentrations were associated with an increased mortality rate, while CO, NO2, and PM10 concentrations had negative associations. In the subgroup analysis, positive associations between SO2 concentrations and IHD mortality were consistently observed in all subgroups, while other pollutant-disease pairs showed null, or mixed associations. Conclusion Long-term exposure to high SO2 concentration was significantly and consistently associated with a high mortality rate nationwide and in Capital and non-Capital areas, and in urban and rural areas. Associations between other air pollutants and disease-related mortalities need to be investigated in further studies.

Abstract Background Emergency department overcrowding negatively impacts critically ill patients and could lead to the occurrence of cardiac arrest. However, the association between emergency department crowding and the occurrence of in-hospital cardiac arrest has not been thoroughly investigated. This study aimed to evaluate the correlation between emergency department occupancy rates and the incidence of in-hospital cardiac arrest. Methods A single-center, observational, registry-based cohort study was performed including all consecutive adult, non-traumatic in-hospital cardiac arrest patients between January 2014 and June 2017. We used emergency department occupancy rates as a crowding index at the time of presentation of cardiac arrest and at the time of maximum crowding, and the average crowding rate for the duration of emergency department stay for each patient. To calculate incidence rate, we divided the number of arrest cases for each emergency department occupancy period by accumulated time. The primary outcome is the association between the incidence of in-hospital cardiac arrest and emergency department occupancy rates. Results During the study period, 629 adult, non-traumatic cardiac arrest patients were enrolled in our registry. Among these, 187 patients experienced in-hospital cardiac arrest. Overall survival discharge rate was 24.6%, and 20.3% of patients showed favorable neurologic outcomes at discharge. Emergency department occupancy rates were positively correlated with in-hospital cardiac arrest occurrence. Moreover, maximum emergency department occupancy in the critical zone had the strongest positive correlation with in-hospital cardiac arrest occurrence (Spearman rank correlation ρ = 1.0, P

ObjectivesForeign patients are more likely to receive inappropriate health service in the emergency room. This study aimed to investigate whether there is health inequality between foreigners and natives who visited emergency rooms with injuries and to examine its causes. MethodsWe analyzed clinical data from the National Emergency Department Information System database associated with patients of all age groups visiting the emergency room from 2013 to 2015. We analyzed data regarding mortality, intensive care unit admission, emergency operation, severity, area, and transfer ratio.ResultsA total of 4,464,603 cases of injured patients were included, of whom 67,683 were foreign. Injury cases per 100,000 population per year were 2,960.5 for native patients and 1,659.8 for foreign patients. Foreigners were more likely to have no insurance (3.1% vs. 32.0%, p < 0.001). Serious outcomes (intensive care unit admission, emergency operation, or death) were more frequent among foreigners. In rural areas, the difference between serious outcomes for foreigners compared to natives was greater (3.7% for natives vs. 5.0% for foreigners, p < 0.001). The adjusted odds ratio for serious outcomes for foreign nationals was 1.412 (95% confidence interval [CI], 1.336–1.492), and that for lack of insurance was 1.354 (95% CI, 1.314–1.394).ConclusionsInjured foreigners might more frequently suffer serious outcomes, and health inequality was greater in rural areas than in urban areas. Foreign nationality itself and lack of insurance could adversely affect medical outcomes.

Background/purposeThe use of MRI as a diagnostic tool has gained popularity in the field of orthopedics. Although 3-dimensional (3D) MRI offers more intuitive visualization and can better facilitate treatment planning than 2-dimensional (2D) MRI, manual segmentation for 3D visualization is time-consuming and lacks reproducibility. Recent advancements in deep learning may provide a solution to this problem through the process of automatic segmentation. The purpose of this study was to develop automated semantic segmentation on 2D MRI images of rotator cuff tears by using a convolutional neural network to visualize 3D models of related anatomic structures.MethodsMRI scans from 56 patients with rotator cuff tears (T2 Linear Coronal MRI; 3.0T, 512 mm × 512 mm, and 2.5-mm slice thickness) were collected. Segmentation masks for the cuff tendon, muscle, bone, and cartilage were obtained by four orthopedic shoulder surgeons, and these data were revised by a shoulder surgeon with more than 20 years' experience. We performed 2D and 3D segmentation using nnU-Net with secondary labels for reducing false positives. Final validation was performed in an external T2 MRI dataset (10 cases) acquired from other institutions. The Dice Similarity Coefficient (DSC) was used to validate segmentation quality.ResultsThe use of 3D nnU-Net with secondary labels to reduce false positives achieved satisfactory results, even with a limited amount of data. The DSCs (mean ± SD) of the cuff tendon, muscle, bone, and cartilage in the internal test set were 80.7% ± 9.7%, 85.8% ± 8.6%, 97.8% ± 0.6%, and 80.8% ± 15.1%, respectively. In external validation, the DSC of the tendon segmentation was 82.74±5.2%.ConclusionAutomated segmentation using 3D U-Net produced acceptable accuracy and reproducibility. This method could provide rapid, intuitive visualization that can significantly facilitate the diagnosis and treatment planning in patients with rotator cuff tears.

The quality of spine surgery planning using X-ray images depends on the accuracy of the delineating landmarks. Owing to the extensive number of landmarks, each planning needs considerable time per patient. This could lead to fatigue and inaccurate planning due to susceptibility to inter- and intra-observer variabilities. Therefore, we propose a fully automated landmark detection in spine X-ray images with a cascaded convolutional neural net (CNN). Five hundred 5K by 2K spine X-ray images from Severance Hospital were acquired. The dataset was split into training, validation, and test sets at a ratio of 8:1:1. The five landmarks for surgery planning from each image were identified by an expert neurosurgeon. To improve accuracy, a two-step cascaded CNN was used. In the first step, the regions of interest (ROIs) were predicted using RetinaNet with focal loss. At this stage, we conducted experiments in various ROI sizes to identify the optimal ROI size for each landmark. In the second step, U-Net predicted the landmarks in the ROIs precisely, and experiments were conducted on different mask sizes to make more accurate predictions; the optimal mask size is presented for each landmark. The errors (mean ± SD) of ROI detection-only model and segmentation-only model were 3.61 ± 3.67 and 7.01 ± 12.53 mm, respectively. The errors of the cascaded CNN models with fixed size and optimal size were 2.58 ± 1.92 mm and 2.08 ± 1.33 mm, respectively. The cascaded semantic segmentation was developed and evaluated for determining landmarks for surgery planning in spine X-ray images, which could be used in clinical settings to reduce the clinical burden of spine surgery planning.

Precision medicine is the personalization of medicine to suit a specific group of people or even an individual patient, based on genetic or molecular profiling. This can be done using genomic, transcriptomic, epigenomic or proteomic information. Personalized medicine holds great promise, especially in cancer therapy and control, where precision oncology would allow medical practitioners to use this information to optimize the treatment of a patient. Personalized oncology for groups of individuals would also allow for the use of population group specific diagnostic or prognostic biomarkers. Additionally, this information can be used to track the progress of the disease or monitor the response of the patient to treatment. This can be used to establish the molecular basis for drug resistance and allow the targeting of the genes or pathways responsible for drug resistance. Personalized medicine requires the use of large data sets, which must be processed and analysed in order to identify the particular molecular patterns that can inform the decisions required for personalized care. However, the analysis of these large data sets is difficult and time consuming. This is further compounded by the increasing size of these datasets due to technologies such as next generation sequencing (NGS). These difficulties can be met through the use of artificial intelligence (AI) and machine learning (ML). These computational tools use specific neural networks, learning methods, decision making tools and algorithms to construct and improve on models for the analysis of different types of large data sets. These tools can also be used to answer specific questions. Artificial intelligence can also be used to predict the effects of genetic changes on protein structure and therefore function. This review will discuss the current state of the application of AI to omics data, specifically genomic data, and how this is applied to the development of personalized or precision medicine on the treatment of cancer.

The artificial neural network (ANN), one of the machine learning (ML) algorithms, inspired by the human brain system, was developed by connecting layers with artificial neurons. However, due to the low computing power and insufficient learnable data, ANN has suffered from overfitting and vanishing gradient problems for training deep networks. The advancement of computing power with graphics processing units and the availability of large data acquisition, deep neural network outperforms human or other ML capabilities in computer vision and speech recognition tasks. These potentials are recently applied to healthcare problems, including computer-aided detection/diagnosis, disease prediction, image segmentation, image generation, etc. In this review article, we will explain the history, development, and applications in medical imaging