Search

Efficient and adaptive computer vision systems have been proposed to make computer vision tasks, such as image classification and object detection, optimized for embedded or mobile devices. These solutions, quite recent in their origin, focus on optimizing the model (a deep neural network) or the system by designing an adaptive system with approximation knobs. Despite several recent efforts, we show that existing solutions suffer from two major drawbacks. First , while mobile devices or systems-on-chips usually come with limited resources including battery power, most systems do not consider the energy consumption of the models during inference. Second , they do not consider the interplay between the three metrics of interest in their configurations, namely, latency, accuracy, and energy. In this work, we propose an efficient and adaptive video object detection system— Virtuoso , which is jointly optimized for accuracy, energy efficiency, and latency. Underlying Virtuoso is a multi-branch execution kernel that is capable of running at different operating points in the accuracy-energy-latency axes, and a lightweight runtime scheduler to select the best fit execution branch to satisfy the user requirement. We position this work as a first step in understanding the suitability of various object detection kernels on embedded boards in the accuracy-latency-energy axes, opening the door for further development in solutions customized to embedded systems and for benchmarking such solutions. Virtuoso is able to achieve up to 286 FPS on the NVIDIA Jetson AGX Xavier board, which is up to 45× faster than the baseline EfficientDet D3 and 15× faster than the baseline EfficientDet D0. In addition, we also observe up to 97.2% energy reduction using Virtuoso compared to the baseline YOLO (v3)—a widely used object detector designed for mobiles. To fairly compare with Virtuoso , we benchmark 15 state-of-the-art or widely used protocols, including Faster R-CNN (FRCNN) [NeurIPS’15], YOLO v3 [CVPR’16], SSD [ECCV’16], EfficientDet [CVPR’20], SELSA [ICCV’19], MEGA [CVPR’20], REPP [IROS’20], FastAdapt [EMDL’21], and our in-house adaptive variants of FRCNN+, YOLO+, SSD+, and EfficientDet+ (our variants have enhanced efficiency for mobiles). With this comprehensive benchmark, Virtuoso has shown superiority to all the above protocols, leading the accuracy frontier at every efficiency level on NVIDIA Jetson mobile GPUs. Specifically, Virtuoso has achieved an accuracy of 63.9%, which is more than 10% higher than some of the popular object detection models, FRCNN at 51.1% and YOLO at 49.5%.

The combination of hyperspectral imaging with multivariate data analysis methods has recently been applied to develop a nondestructive technique, required to determine the seed viability of artificially aged vegetable and cereal seeds. In this study, the potential of shortwave infrared hyperspectral imaging to determine the viability of naturally aged seeds was investigated and thereafter a model for online seed sorting system was developed. The hyperspectral images of 400 Hinoki cypress tree seeds were acquired, and germination tests were conducted for viability confirmation, which indicated 31.5% of the viable seeds. Partial least square discriminant analysis models with 179 variables in the wavelength region of 1000–1800 nm were developed with a maximum model accuracy of 98.4% and 93.8% in both the calibration and validation sets, respectively. The partial least square discriminant analysis beta coefficient revealed the key wavelengths to differentiate viable from nonviable seeds, determined based on the differences in the chemical compositions of the seeds, including their lipid and fatty acid contents, which may control the germination ability of the seeds. The most effective wavelengths were selected using two model-based variable selection methods (i.e., the variable importance of projection (15 variables) and the successive projections algorithm (8 variables)) to develop the model. The successive projections algorithm wavelength selection method was considered to develop a viability model, and its application to the raw data resulted in a prediction accuracy of 94.7% in the calibration set and 92.2% in the validation set. These results demonstrate the potential of shortwave infrared hyperspectral imaging spectroscopy as a powerful nondestructive method to determine the viability of Hinoki cypress seeds. This method could be applied to develop an online seed sorting system for seed companies and nurseries.

We evaluated whether there is a difference in the local recurrence and survival after pelvic external radiotherapy (ERT) with and without boost vaginal brachytherapy (VB) in cervical cancer patients with positive or close vaginal resected margins (RM). We retrospectively reviewed FIGO stage IA-IIB cervical cancer patients treated with postoperative ERT between 1997 and 2018. The sixty patients showing close (safety margin

Background: The purpose of this study was to examine the mediating effects of achievement goals (performance approach goal, mastery approach goal, performance avoidance goal) on the link between role stress and burnout and engagement. Methods: A total 224 non-commissioned officers agreed to the purpose of this study were participated. A correlation analysis was performed by using SPSS 21.0, and Structural Equation Modeling (SEM) was performed by using AMOS 20.0. Results: This study were as follows. As a result of comparing the research model including the performance approach goal and the competition model except the performance approach goal, the competition model was excellent. In addition, performance avoidance goal was partially mediated in the relationship between role stress and burnout. In addition, the mastery approach goal was partially in the relationship between role stress and engagement. Conclusions: The results imply that improving mastery approach goal and decreasing performance avoidance goal is very important for non-commissioned officers. Therefore, the military authorities should develop various strategies to decrease performance avoidance goal and improve the mastery approach goal.

Problems caused by unapproved living modified organism (LMO) seeds in seed imports have emphasized determination of sample size as the most fundamental step for detecting them. Even though a guideline exists, no studies have considered the weight of seeds or validated the sample size through experiments. Hence, this study aimed at determining the optimal sample size for detecting LMO seeds by considering weights of various seeds and validating this via experiment. Binomial distribution was employed to determine the sample size for 26 kinds of seeds, and the size was validated by using 0.1% LMO seed adulteration for small and large seed types. The results indicated that the optimal sample size was more than 50, 500, and 1300 g for small, medium, and large seeds, respectively, at the 99% confidence level. At the 95% confidence level, the optimal sample size was determined as 20, 300, and 900 g, respectively. The results demonstrated 100% detection ability for both large and small seed types, validating the determined sample size. Confidence level and degree of LMO adulteration, rather than population size, were revealed as the factors determining sample size. In conclusion, this study determined and validated the effective sample size for LMO seed detection, which is highly applicable in the field of seed inspection and for quarantine agencies.

The seeds of unapproved living modified organism (LMO) have caused environmental problems; therefore, there has been emphasis on intensive quarantine before complete seed import. For developing a proper sampling method, statistical analysis has been applied, but the current methods generally assume a homogeneous seed mixture and none of the studies have investigated a sampling method for the non-homogeneous case. Therefore, the aim of this study was to determine an optimal sampling method applicable for an inhomogeneous mixture of LMO and normal seeds. By using both statistical and empirical approaches, we demonstrated that sampling on stratified and systemically compartmentalized positions with a unit size was the best option for the inhomogeneous case. Our results are expected to aid in developing a sampling strategy for practical use in quarantine agency.

Adaptive and efficient computer vision systems have been proposed to make computer vision tasks, e.g., object classification and object detection, optimized for embedded boards or mobile devices. These studies focus on optimizing the model (deep network) or system itself, by designing an efficient network architecture or adapting the network architecture at runtime using approximation knobs, such as image size, type of object tracker, head of the object detector (e.g., lighter-weight heads such as one-shot object detectors like YOLO over two-shot object detectors like FRCNN). In this work, we benchmark different video object detection protocols, including FastAdapt, with respect to accuracy, latency, and energy consumption on three different embedded boards that represent the leading edge mobile GPUs. Our set of protocols consists of Faster R-CNN, YOLOv3, SELSA, MEGA, and REPP. Further, we characterize their performance under different levels of resource contention, specifically GPU contention, as would arise due to co-located applications on these boards, contending with the video object detection task. Our key insights are that object detectors have to be coupled with trackers to keep up with the latency requirements (e.g., 30 fps). With this, FastAdapt achieves up to 76 fps on the most well-resourced NVIDIA Jetson-class board---the NVIDIA AGX Xavier. Second, adaptive protocols like FastAdapt, FRCNN, and YOLO (specifically our adaptive variants, FRCNN+ and YOLO+) work well under resource constraints. Among the latest video object detection heads, SELSA achieves the highest accuracy but at a latency of over 2 sec per frame. Our energy consumption experiments bring out that FastAdapt, adaptive FRCNN, and adaptive YOLO are best-in-class, relative to the non-adaptive protocols SELSA, MEGA, and REPP.

BACKGROUND In the era of artificial intelligence, event prediction models are abundant. However, considering the limitation of the electronic medical record–based model, including the temporally skewed prediction and the record itself, these models could be delayed or could yield errors. OBJECTIVE In this study, we aim to develop multiple event prediction models in intensive care units to overcome their temporal skewness and evaluate their robustness against delayed and erroneous input. METHODS A total of 21,738 patients were included in the development cohort. Three events—death, sepsis, and acute kidney injury—were predicted. To overcome the temporal skewness, we developed three models for each event, which predicted the events in advance of three prespecified timepoints. Additionally, to evaluate the robustness against input error and delays, we added simulated errors and delayed input and calculated changes in the area under the receiver operating characteristic curve (AUROC) values. RESULTS Most of the AUROC and area under the precision-recall curve values of each model were higher than those of the conventional scores, as well as other machine learning models previously used. In the error input experiment, except for our proposed model, an increase in the noise added to the model lowered the resulting AUROC value. However, the delayed input did not show the performance decreased in this experiment. CONCLUSIONS For a prediction model that was applicable in the real world, we considered not only performance but also temporal skewness, delayed input, and input error.

PURPOSE To evaluate the treatment outcomes of adjuvant external beam radiation therapy (EBRT) and vaginal brachytherapy (VB) following radical hysterectomy in cervical cancer patients with involved vaginal resection margin (VRM). Materials and. METHODS We retrospectively reviewed the medical records of 21 patients treated with postoperative EBRT and VB for positive VRM FIGO stage IB-IIA cervical cancer between 2003 and 2015. Concurrent platinum-based chemotherapy was administered to all patients. RESULTS The median whole pelvis EBRT dose was 50.4 Gy (range, 45 to 50.4 Gy). In the VB, the median dose per fraction, number of fractions, and total dose delivered were: 4 Gy (range, 3.0 to 4.0 Gy), 4 fractions (range, 3 to 5 fractions), and 16 Gy (range, 12 to 20 Gy), respectively. At a median follow-up of 46 months (range, 9 to 122 months), local recurrence was observed in 2 patients, and distant metastasis was present in 7 patients. All patients with local recurrence subsequently developed distant metastases. The 5-year local control, disease-free survival, and overall survival rates were 89.1%, 65.9%, and 62.9%, respectively. Of the 21 patients, 7 patients (33.3%) reported grade 2 acute toxicity; however, there were no grade 3 or higher acute adverse events. Grade 1-2 late toxicities were observed in 8 patients. Late grade 3 urinary toxicity was reported in 1 patient. Conclusions: Adjuvant EBRT and VB showed excellent local control and low toxicity in cervical cancer patients with positive VRM. Although limited by its retrospective nature, the findings from our study provide evidence supporting the use of additional VB in pathologically involved VRM.

In-vivo dosimetry, in which small detector elements are attached to a patient’s body, is an important technique for directly evaluating radiation treatment doses. The glass rod dosimeter (GRD) possesses several advantages over alternatives, which makes it one of the most useful detectors for in-vivo dosimetry. However, because the GRD initialization process requires a prolonged exposure at very high temperatures, as well as subsequent gradual quenching, each measurement takes approximately a day to complete. Therefore, we investigated the reliability of a GRD used repeatedly without initialization processes to improve efficiency. Ten doses of 0.5 Gy were delivered and read using three GRD elements. Then, the same procedure was performed for doses of 1.0 Gy. A readout error of less than 2% was maintained for up to three irradiation doses. However, the fluctuations in the readout data increased significantly as the number of irradiation doses increased. In addition, we discovered that the combined uncertainty of the readouts was influenced more heavily by the cumulative amount of irradiation than it was by the number of doses. Our results should provide guidance for accurate and efficient GRD use.

The X-ray imaging modality has played an important role in the daily verification of a patient’s position for accurate radiation treatment (RT), especially in modern RT techniques that require highly precise patient localization. Although cone-beam computer tomography (CBCT) has already been introduced to the RT field, two-dimensional (2D) X-ray portal imagers have been more widely used than CBCT for the daily check of patient localization owing to the lower patient exposure and time consumption. However, the 2D imager practically provide a lower bone-tissue contrast than CBCT. Thus, we propose a method that enhances the image contrast of daily acquired 2D images by using just one CBCT scan during an entire RT course. This method operates on the basis of the depth normalization (DN) technique and requires no artificial data manipulation such as image post-processing filters. We implemented the algorithm for the portal images of a cylindrical phantom and three patients. From DN results, the image contrast for the phantom and the patients increased by factors of 26.3 and 13.4 on average, and their contrast-to-noise ratios were maintained at differences of 8.6% and 7.7% on average, respectively. Moreover, the DN method provided a stronger contrast enhancement at lower doses that could suppress the imaging exposure and thus improve the patient’s safety. Therefore, the verification of RT patient localization is expected to be performed more efficiently and accurately by using the DN method.

Primary liver tumor, especially hepatocellular carcinoma (HCC), is a common cause of cancer death worldwide. The incidence is generally higher in Asian countries than in western countries. Carcinogenesis of HCC is often associated with hepatitis viral infections. Current standard treatment of HCC is surgical resection or transplantation in patients with early stage disease. However, the patient with advanced stage disease, surgical resection is often limited. Sorafenib or other treatment modalities are not so effective as well. We report a case of unusual radiation super-sensitivity in advanced stage HCC, and review the literature.

The kidney is one of the most radiosensitive organs in the abdominal cavity and is the dose-limiting structure in cancer patients receiving abdominal or total body irradiation. In the present study, the effect of coenzyme Q10 (CoQ10) on radiation nephropathy was evaluated in rats. A total of 72 rats were equally randomized into 4 groups: Control, CoQ10, irradiation with 10 Gy (RT) + placebo, or RT + CoQ10. The 2 RT groups received single 10 Gy of abdominal irradiation. The 2 CoQ10 groups were supplemented daily with 1 mL of soybean oil containing 10 mg/kg of CoQ10. The RT + placebo and control groups received same dose of soybean oil. After 24 weeks, laboratory and histopathologic findings were compared. The 2 RT groups showed significant increases in blood urea nitrogen (BUN) and creatinine levels and significant pathologic changes such as glomerulosclerosis and tubulointerstitial fibrosis. CoQ10 supplementation resulted in significant reductions of BUN and creatinine levels compared with the RT + placebo group (P < 0.001 and P = 0.038, respectively). CoQ10 treatment significantly attenuated glomerular and tubular changes of irradiated kidney in semiquantitative analysis (P < 0.001 for both). Administration of CoQ10 can alleviate the radiation-induced nephropathy., Graphical Abstract

The grading of ginseng (Panax ginseng) including the evaluation of internal quality attributes is essential in the ginseng industry for quality control. Assessment for inner whitening, a major internal disorder, must be conducted when identifying high quality ginseng. Conventional methods for detecting inner whitening in ginseng root samples use manual inspection, which is time-consuming and inaccurate. This study develops an internal quality measurement technique using near-infrared transmittance spectral imaging to evaluate inner whitening in ginseng samples. Principle component analysis (PCA) was used on ginseng hypercube data to evaluate the developed technique. The transmittance spectra and spectral images of ginseng samples exhibiting inner whitening showed weak intensity characteristics compared to normal ginseng in the region of 900&ndash, 1050 nm and 1150&ndash, 1400 nm respectively, owing to the presence of whitish internal tissues that have higher optical density. On the basis of the multivariate analysis method, even a simple waveband ratio image has the great potential to quickly detect inner whitening in ginseng samples, since these ratio images show a significant difference between whitened and non-whitened regions. Therefore, it is possible to develop an efficient and rapid spectral imaging system for the real-time detection of inner whitening in ginseng using minimal spectral wavebands. This novel strategy for the rapid, cost-effective, non-destructive detection of ginseng&rsquo, s inner quality can be a key component for the automation of ginseng grading.

Mandarin orange is a popularly consumed fruit in Asian countries. Over 99% of cultivation area in Korea for mandarin oranges is concentrated in Jeju Island. Despite of this high concentration, detecting infection and estimating fruit yields has been done manually, resulting in loss of money and time. In this study, hyperspectral fluorescence imaging technique was explored to distinguish green mandarin oranges from leaves to estimate fruit density. In addition, early stage detection for disease infection of leaves and fruits were investigated. The fluorescence spectral images showed reliable performance for distinguishing green mandarin oranges from leaves, and detecting disease infection on both leaves and fruits. The result demonstrated that hyperspectral fluorescence imaging might be used for rapid and non-destructive detection of disease infection and yield estimation of mandarin orange in the field.

Purpose: The overall goal of this study is to restore kilovoltage computed tomography (kV-CT) images which are disfigured by patients’ metal prostheses. By generating a hybrid sinogram that is a combination of kV and megavoltage (MV) projection data, the authors suggest a novel metal artifact-reduction (MAR) method that retains the image quality to match that of kV-CT and simultaneously restores the information of metal prostheses lost due to photon starvation. Methods: CT projection data contain information about attenuation coefficients and the total length of the attenuation. By normalizing raw kV projections with their own total lengths of attenuation, mean attenuation projections were obtained. In the same manner, mean density projections of MV-CT were obtained by the normalization of MV projections resulting from the forward projection of density-calibrated MV-CT images with the geometric parameters of the kV-CT device. To generate the hybrid sinogram, metal-affected signals of the kV sinogram were identified and replaced by the corresponding signals of the MV sinogram following a density calibration step with kV data. Filtered backprojection was implemented to reconstruct the hybrid CT image. To validate the authors’ approach, they simulated four different scenarios for three heads and one pelvis using metallic rod inserts within a cylindrical phantom. Five inserts describing human body elements were also included in the phantom. The authors compared the image qualities among the kV, MV, and hybrid CT images by measuring the contrast-to-noise ratio (CNR), the signal-to-noise ratio (SNR), the densities of all inserts, and the spatial resolution. In addition, the MAR performance was compared among three existing MAR methods and the authors’ hybrid method. Finally, for clinical trials, the authors produced hybrid images of three patients having dental metal prostheses to compare their MAR performances with those of the kV, MV, and three existing MAR methods. Results: The authors compared the image quality and MAR performance of the hybrid method with those of other imaging modalities and the three MAR methods, respectively. The total measured mean of the CNR (SNR) values for the nonmetal inserts was determined to be 14.3 (35.3), 15.3 (37.8), and 25.5 (64.3) for the kV, MV, and hybrid images, respectively, and the spatial resolutions of the hybrid images were similar to those of the kV images. The measured densities of the metal and nonmetal inserts in the hybrid images were in good agreement with their true densities, except in cases of extremely low densities, such as air and lung. Using the hybrid method, major streak artifacts were suitably removed and no secondary artifacts were introduced in the resultant image. In clinical trials, the authors verified that kV and MV projections were successfully combined and turned into the resultant hybrid image with high image contrast, accurate metal information, and few metal artifacts. The hybrid method also outperformed the three existing MAR methods with regard to metal information restoration and secondary artifact prevention. Conclusions: The authors have shown that the hybrid method can restore the overall image quality of kV-CT disfigured by severe metal artifacts and restore the information of metal prostheses lost due to photon starvation. The hybrid images may allow for the improved delineation of structures of interest and accurate dose calculations for radiation treatment planning for patients with metal prostheses.

One of the emerging issues in radiography is low-dose imaging to minimize patient’s exposure. The scintillating materials employed in most indirect flat-panel detectors show a drastic change of X-ray photon absorption efficiency around their K-edge energies that consequently affects image quality. Using various tube voltages, we investigated the imaging performance of most popular scintillators: cesium iodide (CsI) and gadolinium oxysulfide (Gd2O2S). The integrated detective quantum efficiencies (iDQE) of four detectors installed in the same hospital were evaluated according to the standardized procedure IEC 62220-1 at tube voltages of 40 - 120 kVp. The iDQE values of the Gd2O2S detectors were normalized by those of CsI detectors to exclude the effects of image postprocessing. The contrast-to-noise ratios (CNR) were also evaluated by using an anthropomorphic chest phantom. The iDQE of the CsI detector outperformed that of the Gd2O2S detector over all tube voltages. Moreover, we noted that the iDQE of the Gd2O2S detectors quickly rolled off with decreasing tube voltage under 70 kVp. The CNRs of the two scintillators were similar at 120 kVp. At 60 kVp, however, the CNR of Gd2O2S was about half that of CsI. Compared to the Gd2O2S detectors, variations in the DQE performance of the CsI detectors were relatively immune to variations in the applied tube voltages. Therefore, we claim that Gd2O2S detectors are inappropriate for use in low-tube-voltage imaging (e.g., extremities and pediatrics) with low patient exposure.

Accurate dose delivery is crucial to the success of modern radiotherapy. To evaluate the dose actually delivered to patients, in-vivo dosimetry (IVD) is generally performed during radiotherapy to measure the entrance doses. In IVD, a build-up device should be placed on top of an in-vivo dosimeter to satisfy the electron equilibrium condition. However, a build-up device made of tissue-equivalent material or metal may perturb dose delivery to a patient, and requires an additional laborious and time-consuming process. We developed a novel IVD method using a look-up table of conversion ratios instead of a build-up device. We validated this method through a monte-carlo simulation and 31 clinical trials. The mean error of clinical IVD is 3.17% (standard deviation: 2.58%), which is comparable to that of conventional IVD methods. Moreover, the required time was greatly reduced so that the efficiency of IVD could be improved for both patients and therapists.

This paper presents the NDIR CO2 gas sensor that has improved the sensitivity and also the accuracy by the deposition of thin hydrophobic film (Parylene-C film with 0.5 micrometer thick) onto the reflector surfaces of White-cell structure. After deposition of hydrophobic thin film, the sensitivity of sensor has been increased with averaged 10% and the estimated errors were reduced within 13 ppm to 143 ppm from 254 K to 324 K temperature ranges and from 0 ppm to 5000 ppm CO2 concentrations.

Fishes are a widely used food material in the world. Recently about 4% of the fishes are infected with Kudoa thyrsites in Asian ocean. Kudoa thyrsites is a parasite that is found within the muscle fibers of fishes. The infected fishes can be a reason of food poisoning, which should be sorted out before distribution and consumption. Although Kudoa thyrsites is visible to the naked eye, it could be easily overlooked due to the micro-scale size and similar color with fish tissue. In addition, the visual inspection is labor intensive works resulting in loss of money and time. In this study, a portable microscopic camera was utilized to obtain images of raw fish slices. The optimized image processing techniques with polarized transmittance images provided reliable performance. The result shows that the portable microscopic imaging method can be used to detect parasites rapidly and non-destructively, which could be an alternative to manual inspections.

Definitive chemoradiotherapy (CRT) followed by brachytherapy is a standard treatment for locally advanced cervical cancer. During CRT, marked reduction of cervical tumor is often observed in magnetic resonance imaging (MRI). The primary aim of this study was to assess the association between tumor response in MRI using FIGO classification and clinical outcomes.Multi-institutional data were retrospectively reviewed to identify the significance of MR tumor response on tumor recurrence and patient survival. 225 patients with histologically confirmed squamous cell carcinoma of the cervix, staged as FIGO Ib2-IVa on initial pelvic MRI, were included. Post-CRT MRI was performed median 35days after the beginning of CRT and before brachytherapy. A median 54Gy of external radiation was given with weekly cisplatin during CRT.112 (49.7%) of the 225 patients showed a positive response in post-CRT MRI and were named the responsive arm. After a median follow-up time of 36.2months, the responsive arm had significantly lower para-aortic recurrence (7.5% vs. 12.4%; p=0.04) and distant metastasis (13.2% vs. 27.6%; p=0.03) rates than did the non-responsive arm. The responsive arm had significantly higher 3-year cause-specific survival rate (94.6% vs. 81.1%, p0.01) than did the non-responsive arm. In the multivariate analysis, tumor size (hazard ratio, 1.91 and 95% confidence interval, 1.07-3.43; p=0.028) and positive MR response (hazard ratio, 1.75 and 95% confidence interval, 1.06-2.27; p=0.045) were significant factors for recurrence-free survival CONCLUSION: Early tumor response evaluation with MRI using FIGO classification effectively predicted distant tumor metastasis and disease-specific survival in locally advanced cervical cancer.