Search

Locating lung field is a critical and fundamental processing stage in the automated analysis of chest radiographs (CXRs) for pulmonary disorders. During the routine examination of CXRs, using both frontal and lateral CXRs can benefit clinical diagnosis of cardiothoracic and lung diseases. However, the accurate segmentation of lung fields on both frontal and lateral CXRs is still challenging due to the blurry boundary of the lung field on lateral CXRs and the poor generalization ability of the models. Existing deep learning-based methods focused on lung field segmentation on frontal CXRs, and the generalization ability of these methods on the different type of CXRs (e.g., pediatric CXRs) and new lung diseases (e.g., COVID-19) has not been tested. In this paper, a view identification assisted fully convolutional network (VI-FCN) is proposed for the segmentation of lung fields on frontal and lateral CXRs simultaneously. The VI-FCN consists of an FCN branch for lung field segmentation and a view identification branch for identification of the frontal and lateral CXRs and for enhancing the lung field segmentation. To improve the generalization ability of VI-FCN, six public datasets and our frontal and lateral CXRs (over 2000 CXRs) were collected for training. The segmentation of lung fields on the Japanese Society of Radiological Technology (JSRT) dataset yields mean dice similarity coefficient (DSC) of 0.979 ± 0.008, mean Jaccard index ( Ω) of 0.959 ± 0.016, and mean boundary distance (MBD) of 1.023 ± 0.487 mm. Besides, the VI-FCN achieves mean DSC of 0.973 ± 0.010, mean Ω of 0.947 ± 0.018, and mean MBD of 1.923 ± 0.755 mm for the segmentation of lung fields on our lateral CXRs. The experiments demonstrate the superior performance of the proposed VI-FCN over most of existing state-of-the-art methods. Moreover, the proposed VI-FCN achieves promising results on untrained pediatric CXRs and COVID-19 datasets.

Objective: accurately classifying the malignancy of lesions detected in a screening scan is critical for reducing false positives. Radiomics holds great potential to differentiate malignant from benign tumors by extracting and analyzing a large number of quantitative image features. Since not all radiomic features contribute to an effective classifying model, selecting an optimal feature subset is critical. Methods: this work proposes a new multi-objective based feature selection (MO-FS) algorithm that considers sensitivity and specificity simultaneously as the objective functions during feature selection. For MO-FS, we developed a modified entropy-based termination criterion that stops the algorithm automatically rather than relying on a preset number of generations. We also designed a solution selection methodology for multi-objective learning that uses the evidential reasoning approach (SMOLER) to automatically select the optimal solution from the Pareto-optimal set. Furthermore, we developed an adaptive mutation operation to generate the mutation probability in MO-FS automatically. Results: we evaluated the MO-FS for classifying lung nodule malignancy in low-dose CT and breast lesion malignancy in digital breast tomosynthesis. Conclusion: the experimental results demonstrated that the feature set selected by MO-FS achieved better classification performance than features selected by other commonly used methods. Significance: the proposed method is general and more effective radiomic feature selection strategy.

Automatically matching corresponding regions of interest (ROIs) on two-view images is valuable in breast cancer diagnosis, benefiting of saving time and cutting the workload. We propose a method for matching the corresponding ROIs by integrating the geometric model and image similarity searching. The geometric model is implemented by restoring a free breast in the 3D space from two-view preprocessed breast contours. Then, the possible position of the ROI center on cranio-caudal (CC)/medio-lateral oblique (MLO) view image is represented by three feature points in the 3D space. As the view changes, these points can be mapped onto the MLO/CC view image. A matching strip is created later according to the confidence interval, within which the specific position of the ROI can be located by image similarity searching. The experiments were conducted on 273 pairs of mammograms with 400 calcifications and 284 pairs with 300 masses to verify the accuracy of the geometric model and similarity searching. The mean absolute error between the curves and the ROI centers was 3.36 ± 2.90 mm. For 95% detection sensitivity, the confidence interval was ±8.77 mm. For calcifications, the mean distance between the centers of the matched ROIs and the reference was 3.92 ± 4.61 mm. About 93.46% cases had overlap greater than 50%, and 92.46% cases had overlap greater than 75%. For masses, the mean distance was 6.15 ± 7.08 mm. About 88.46% cases had overlap greater than 50%, and 85.58% cases had overlap greater than 75%.

Bone suppression of chest radiographs (CXRs) is potentially useful for diagnosing lung diseases for radiologists and computer–aided diagnosis. This paper presents a cascaded convolutional network model in wavelet domain (Wavelet-CCN) for bone suppression in single conventional CXR. Wavelet coefficients are sparse and suitable as the output of convolutional network. The convolutional networks are trained to predict the wavelet coefficients of bone images from the wavelet coefficients of CXRs, using real two-exposure dual energy subtraction (DES) CXRs as training data. By combining the multi-level wavelet decomposition and a cascaded refinement framework, the Wavelet-CCN model can work automatically with a multi-scale approach and progressively refine the prediction in terms of accuracy and spatial resolution. Compared with previous work of CamsNet model which preforms bone prediction in gradient domain, the Wavelet-CCN model predicts the wavelet coefficients to reconstruct bone images and can avoid the inconsistent background intensity caused by 2D integration of gradients. The predicted bone image is subtracted from the original CXR to produce a soft-tissue image. The Wavelet-CCN model and its variants with different wavelet basis are evaluated on a dataset that consists of 504 cases of real two-exposure DES CXRs (404 cases for training and 100 cases for test). Experimental results show that among all the variants and different wavelet bases, the Wavelet-CCN model with Haar wavelet performs best. The average peak signal-to-noise ratio and structural similarity index of the soft-tissue images produced by the proposed Wavelet-CCN model are both higher than those of the previous CamsNet model in gradient domain, reaching values of 39.4 (±0.94) dB and 0.977 (±0.004), respectively. The results also demonstrate that the Wavelet-CCN model can process the CXRs acquired by four types of X-ray machines.

Objectives. To build and validate random forest (RF) models for the classification of bone tumors based on the conventional radiographic features of the lesion and patients’ clinical characteristics, and identify the most essential features for the classification of bone tumors. Materials and Methods. In this retrospective study, 796 patients (benign bone tumors: 412 cases, malignant bone tumors: 215 cases, intermediate bone tumors: 169 cases) with pathologically confirmed bone tumors from Nanfang Hospital of Southern Medical University, Foshan Hospital of TCM, and University of Hong Kong-Shenzhen Hospital were enrolled. RF models were built to classify tumors as benign, malignant, or intermediate based on conventional radiographic features and potentially relevant clinical characteristics extracted by three musculoskeletal radiologists with ten years of experience. SHapley Additive exPlanations (SHAP) was used to identify the most essential features for the classification of bone tumors. The diagnostic performance of the RF models was quantified using receiver operating characteristic (ROC) curves. Results. The features extracted by the three radiologists had a satisfactory agreement and the minimum intraclass correlation coefficient (ICC) was 0.761 (CI: 0.686-0.824, P < .001 ). The binary and tertiary models were built to classify tumors as benign, malignant, or intermediate based on the imaging and clinical features from 627 and 796 patients. The AUC of the binary (19 variables) and tertiary (22 variables) models were 0.97 and 0.94, respectively. The accuracy of binary and tertiary models were 94.71% and 82.77%, respectively. In descending order, the most important features influencing classification in the binary model were margin, cortex involvement, and the pattern of bone destruction, and the most important features in the tertiary model were margin, high-density components, and cortex involvement. Conclusions. This study developed interpretable models to classify bone tumors with great performance. These should allow radiographers to identify imaging features that are important for the classification of bone tumors in the clinical setting.

This study established an interpretable machine learning model to predict the severity of coronavirus disease 2019 (COVID-19) and output the most crucial deterioration factors. Clinical information, laboratory tests, and chest computed tomography (CT) scans at admission were collected. Two experienced radiologists reviewed the scans for the patterns, distribution, and CT scores of lung abnormalities. Six machine learning models were established to predict the severity of COVID-19. After parameter tuning and performance comparison, the optimal model was explained using Shapley Additive explanations to output the crucial factors. This study enrolled and classified 198 patients into mild ( n = 162 ; 46.93 ± 14.49 years old) and severe ( n = 36 ; 60.97 ± 15.91 years old) groups. The severe group had a higher temperature ( 37.42 ± 0.99 °C vs. 36.75 ± 0.66 °C), CT score at admission, neutrophil count, and neutrophil-to-lymphocyte ratio than the mild group. The XGBoost model ranked first among all models, with an AUC, sensitivity, and specificity of 0.924, 90.91%, and 97.96%, respectively. The early stage of chest CT, total CT score of the percentage of lung involvement, and age were the top three contributors to the prediction of the deterioration of XGBoost. A higher total score on chest CT had a more significant impact on the prediction. In conclusion, the XGBoost model to predict the severity of COVID-19 achieved excellent performance and output the essential factors in the deterioration process, which may help with early clinical intervention, improve prognosis, and reduce mortality.

To investigate the effect of different breast lesions on exposure parameters in digital mammography and to determine whether the exposure parameters can additively improve diagnostic efficiency. Craniocaudal view and mediolateral view full-field digital mammography images from 982 women with unilateral lesions (341 with malignant lesions, 189 with benign lesions, and 452 healthy women) obtained at Nanfang Hospital were reviewed. Differences in exposure parameters (tube voltage and load, breast thickness (BT), and average glandular dose (AGD)) between breasts were calculated. The relationships between parameter differences and lesion size were explored. A logistic regression model was used based on the AGD and BT differences, and the area under the receiver operating characteristic curve (AUC) was used to assess the performance of these parameters in differentiating malignant from benign and healthy subjects. Independently, data from 129 women (82 with malignant and 47 with benign lesions) treated at Sun Yat-sen Memorial Hospital were collected to validate the model. Differences in tube voltage and load, BT, and AGD between breasts were significantly greater in the malignant subjects than benign (p < 0.05) and healthy subjects (p < 0.05). The AUCs for the comparisons of malignant vs. healthy subjects, malignant vs. benign subjects, and benign vs. healthy subjects were 0.77 ± 0.02, 0.72 ± 0.02, and 0.57 ± 0.02, respectively. The model combining the exposure parameters with the BI-RADS category resulted in a higher AUC (0.910 ± 0.03) compared with physician diagnosis alone (0.820 ± 0.04) for differentiating between malignant and benign lesions. Exposure parameters additively improved diagnostic accuracy for breast cancer and yielded more reliable results. • Differences in kVp, mAs, BT, and AGD between breasts were significantly greater in the malignant subjects than benign and healthy subjects. • The model combining exposure parameters with the BI-RADS category resulted in a higher AUC compared with the physician’s diagnosis for differentiating between malignant and benign lesions. • Exposure parameters additively improved diagnostic accuracy for breast cancer.

Objectives. To compare the differences in normalized average glandular dose (NAGD) between the breasts of healthy subjects and those of cancer patients and to determine if the NAGD difference is associated with breast cancer risk and improves breast cancer classification. Materials and Methods. Craniocaudal view and mediolateral view full-field digital mammography (FFDM) images were obtained from 1682 healthy subjects whose breasts were categorized as Breast Imaging-Reporting and Data System (BI-RADS) I or II and from 811 biopsy-confirmed unilateral breast cancer patients whose breasts on the contralateral side were category I or II. Both populations were randomized into training and test sets. Multivariate logistic regression analysis was used to build the breast cancer risk assessment model, and the area under the receiver operating characteristic curve ( A z ) was used to evaluate the model. Twenty-two breast cancer patients who were originally categorized as BI-RADS I or II for both breasts, but were diagnosed with unilateral biopsy-confirmed breast cancer subsequently, were included to validate the model. Results. The NAGD differences in both FFDM images between tumor-bearing breasts and the healthy breasts of patients were significantly higher than those in healthy subjects ( P < 0.001 ). The model with NAGD differences had a higher A z value than the model without NAGD differences. While there was no NAGD differences between originally healthy breasts of breast cancer patients, significant NAGD differences between now tumor-bearing breasts and the then previously healthy breasts were found in both FFDM images. Conclusions. NAGD differences between both breasts can be included in the breast cancer risk assessment model to evaluate breast cancer risk.

Generating dual-energy subtraction (DES) soft-tissue images from chest radiographs (also called bone suppression) is an important task, as it improves the detection rates for lung nodules. Previous studies focus on generating DES-like soft-tissue images from CXRs through machine/deep learning techniques. However, they usually require tedious image processing steps for bone segmentation/delineation or ignore anatomical structure information (e.g., edges of ribs and clavicles) in CXRs. In this work, we propose a bone Edge-guided Generative Adversarial Network (EGAN) to generate DES-like soft-tissue images from conventional CXRs, which does not require human intervention and can explicitly use anatomical structure information of bones in CXRs. Specifically, the edges of ribs and clavicles in an input CXR were first detected by a trained fully convolutional network. Then, the edge probability map, as well as the original CXR image, are fed into a GAN model to generate the DES-like soft-tissue image, where the detected edge information is used as the prior knowledge to directly and specifically guide the image generation process. Experimental results on 504 subjects (each equipped with a CXR, a DES bone image, and a DES soft-tissue image) demonstrate that EGAN can produce DES-like soft-tissue images with high-quality and high-resolution, compared with classic deep learning methods. We further apply the trained EGAN to CXRs acquired by different types of X-ray machines in the public JSRT and NIH ChestXray 14 datasets, and our method can also produce visually appealing DES-like soft-tissue images.

Digital breast tomosynthesis (DBT) with improved lesion conspicuity and characterization has been adopted in screening practice. DBT-based diagnosis strongly depends on physicians' experience, so an automatic lesion malignancy classification model using DBT could improve the consistency of diagnosis among different physicians. Tensor-based approaches that use the original imaging data as input have shown promising results for many classification tasks. However, DBT data are pseudo-3D volumetric imaging as the slice spacing of DBT is much coarser than that of the in-plane resolution. Thus, directly constructing DBT as the third-order tensor in a conventional tensor-based classifier with introducing additional information to the original DBT data along the slice-spacing dimension will lead to inconsistency across all three dimensions. To avoid such inconsistency, we introduce a collection input based support tensor machine (CISTM)-based classifier that uses the tensor collection as input for classifying lesion malignancy in DBT. In CISTM, instead of introducing the third dimension directly into the geometry construction, the third-dimension structural relationship is related by weight parameters in the decision function, which is dynamically and automatically constructed during the classifier training process and is more consistent with the pseudo-3D nature of DBT. We tested our method on a DBT dataset of 926 images among which 262 were malignant and 664 were benign. We compared our method with the latest tensor-based method, KSTM (kernelled support tensor machine), which does not consider the unique non-uniform resolution property of DBT. Experimental results illustrate that the CISTM-based classifier is effective for classifying breast lesion malignancy in DBT and that it outperforms the KSTM-based classifier.

Digital breast tomosynthesis (DBT) mammography is a promising imaging technique for detecting early-stage breast cancers. However, DBT imaging usually contains truncation artifacts around the chest wall area, where many recurring breast tumors or masses are located. Extrapolation or interpolation and weighting techniques are used to suppress artifacts in existing methods, but these are inappropriate for DBT due to its high spatial resolution requirement. To solve this problem, we propose a practical truncation correction method, which can realize simultaneous truncation artifact reduction and image reconstruction (STARIR) for DBT. In contrast to extrapolation or interpolation estimations and weighting strategy, this integrated method renews each truncated voxel by considering the information of all measured projections rather than that of one projection at each reconstruction iteration. Qualitative and quantitative studies were performed on simulated and realistic DBT data to validate the proposed method. Results showed that STARIR reduces truncation artifacts and preserves tissue details in reconstructed images more effectively than do the current methods. The proposed method can be used for simultaneous and accurate truncation artifact reduction and image reconstruction in DBT.

Suppression of bony structures in chest radiographs (CXRs) is potentially useful for radiologists and computer-aided diagnostic schemes. In this paper, we present an effective deep learning method for bone suppression in single conventional CXR using deep convolutional neural networks (ConvNets) as basic prediction units. The deep ConvNets were adapted to learn the mapping between the gradients of the CXRs and the corresponding bone images. We propose a cascade architecture of ConvNets (called CamsNet) to refine progressively the predicted bone gradients in which the ConvNets work at successively increased resolutions. The predicted bone gradients at different scales from the CamsNet are fused in a maximum-a-posteriori framework to produce the final estimation of a bone image. This estimation of a bone image is subtracted from the original CXR to produce a soft-tissue image in which the bone components are eliminated. Our method was evaluated on a dataset that consisted of 504 cases of real two-exposure dual-energy subtraction chest radiographs (404 cases for training and 100 cases for test). The results demonstrate that our method can produce high-quality and high-resolution bone and soft-tissue images. The average relative mean absolute error of the produced bone images and peak signal-to-noise ratio of the produced soft-tissue images were 3.83% and 38.7dB, respectively. The average bone suppression ratio of our method was 83.8% for the CXRs with pixel sizes of nearly 0.194mm. Furthermore, we apply the trained CamsNet model on the CXRs acquired by various types of X-ray machines, including scanned films, and our method can also produce visually appealing bone and soft-tissue images.

Computer aided detection (CADe) for breast lesions can provide an important reference for radiologists in breast cancer screening. Architectural distortion (AD) is a type of breast lesion that is difficult to detect. A majority of CADe methods focus on detecting the radial pattern, which is a main characteristic of typical ADs. However, a few atypical ADs do not exhibit such a pattern. To improve the performance of CADe for typical and atypical ADs, we propose a deep-learning-based model that used mammary gland distribution as prior information to detect ADs in digital breast tomosynthesis (DBT). First, information about gland distribution, including the Gabor magnitude, the Gabor orientation field, and a convergence map, were produced using a bank of Gabor filters and convergence measures. Then, this prior information and an original slice were input into a Faster R-CNN detection network to obtain the 2-D candidates for each slice. Finally, a 3-D aggregation scheme was employed to fuse these 2-D candidates as 3-D candidates for each DBT volume. Retrospectively, 64 typical AD volumes, 74 atypical AD volumes, and 127 normal volumes were collected. Six-fold cross-validation and mean true positive fraction (MTPF) were used to evaluate the model. Compared to an existing convergence-based model, our proposed model achieved an MTPF of 0.53 ± 0.04, 0.61 ± 0.05, and 0.45 ± 0.04 for all DBT volumes, typical + normal volumes, and atypical + normal volumes, respectively. These results were significantly better than those of 0.36 ± 0.03, 0.46 ± 0.04, and 0.28 ± 0.04 for a convergence-based model (p ≪ 0.01). These results indicate that employing the prior information of gland distribution and a deep learning method can improve the performance of CADe for AD.

Objective. To develop and evaluate a multi-path synergic fusion (MSF) deep neural network model for breast mass classification using digital breast tomosynthesis (DBT). Methods. We retrospectively collected 441 patients who had undergone DBT in which the regions of interest (ROIs) covering the malignant/benign breast mass were extracted for model training and validation. In the proposed MSF framework, three multifaceted representations of the breast mass (gross mass, overview, and mass background) are extracted from the ROIs and independently processed by a multi-scale multi-level features enforced DenseNet (MMFED). The three MMFED sub-models are finally fused at the decision level to generate the final prediction. The advantages of the MMFED over the original DenseNet, as well as different fusion strategies embedded in MSF, were comprehensively compared. Results. The MMFED was observed to be superior to the original DenseNet, and multiple channel fusions in the MSF outperformed the single-channel MMFED and double-channel fusion with the best classification scores of area under the receiver operating characteristic (ROC) curve (87.03%), Accuracy (81.29%), Sensitivity (74.57%), and Specificity (84.53%) via the weighted fusion method embedded in MSF. The decision level fusion-based MSF was significantly better (in terms of the ROC curve) than the feature concatenation-based fusion (p< 0.05), the single MMFED using a fused three-channel image (p< 0.04), and the multiple MMFED end-to-end training (p< 0.004). Conclusions. Integrating multifaceted representations of the breast mass tends to increase benign/malignant mass classification performance and the proposed methodology was verified to be a promising tool to assist in clinical breast cancer screening.

Aim To investigate mammography and magnetic resonance imaging (MRI) manifestations of benign and malignant adenomyoepithelioma (AME), in order to improve our understanding of this disease. Materials and methods Thirteen patients (11 cases of benign AME and two cases of malignant AME) were included. All patients underwent preoperative mammography, and four underwent preoperative MRI examinations. Results Mammography revealed that 11 cases showed round, lobulate, or oval masses. One case was accompanied by calcification with blurred edges; the remaining 10 cases showed masses with clear or shaded borders. One case exhibited structurally distorted lesions, and another showed focal opacities. Among the four patients who also underwent preoperative MRI examinations, signals in the three benign cases were clearly enhanced homogeneously, and the dynamic enhancement curves were progressive. The one case of malignant AME showed an irregularly shaped mass with spiculate margins on the MRI images. The signals were homogeneously enhanced, and the dynamic enhancement curve was of the washout type. Conclusion AME usually exhibited oval or round masses with smooth edges and no calcification, and, in a portion of the cases, structurally distorted lesions or focal opacities. On different MRI sequences, AME usually showed homogeneous signal. The apparent diffusion coefficient value of malignant AME is relatively low.

The mammography is the first choice of breast cancer screening, which has proven to be the most effective screening method. An antiscatter grid is usually employed to enhance the contrast of image by absorbing unexpected scattered signals. However, the grid pattern casts shadows and grid artifacts, which severely degrade the image quality. To solve the problem, we propose the patch based frequency signal filtering for fast grid artifacts suppressing. As opposed to whole image processing synchronously, the proposed method divides image into a number of blocks for tuning filter simultaneously, which reduces the frequency interference among image blocks and saves computation time by multithread processing. Moreover, for mitigating grid artifacts more precisely, characteristic peak detection is employed in each block automatically, which can accurately identify the location of the antiscatter grid and its motion pattern. Qualitative and quantitative studies were performed on simulation and real machine data to validate the proposed method. The results show great potential for fast suppressing grid artifacts and generating high quality of digital mammography.

Chest tomosynthesis (CTS) is a newly developed imaging technique which provides pseudo-3D volume anatomical information of thorax from limited angle projections and therefore improves the visibility of anatomy without so much increase on radiation dose compared to the chest radiography (CXR). However, one of the relatively common problems in CTS is the respiratory motion of patient during image acquisition, which negatively impacts the detectability. In this paper, we propose a sin-quadratic model to analyze the respiratory motion during CTS scanning, which is a real time method that generates the respiratory signal by directly extracting the motion of diaphragm during data acquisition. According to the extracted respiratory signal, physicians could re-scan the patient immediately or conduct motion free CTS image reconstruction for patients that could not hold their breath perfectly during the scan time. The effectiveness of the proposed model was demonstrated with both the simulated phantom data and the real patient data.