The Medical Segmentation Decathlon.

International challenges have become the de facto standard for comparative assessment of image analysis algorithms. Although segmentation is the most widely investigated medical image processing task, the various challenges have been organized to focus only on specific clinical tasks. We organized the Medical Segmentation Decathlon (MSD)-a biomedical image analysis challenge, in which algorithms compete in a multitude of both tasks and modalities to investigate the hypothesis that a method capable of performing well on multiple tasks will generalize well to a previously unseen task and potentially outperform a custom-designed solution. MSD results confirmed this hypothesis, moreover, MSD winner continued generalizing well to a wide range of other clinical problems for the next two years. Three main conclusions can be drawn from this study: (1) state-of-the-art image segmentation algorithms generalize well when retrained on unseen tasks; (2) consistent algorithmic performance across multiple tasks is a strong surrogate of algorithmic generalizability; (3) the training of accurate AI segmentation models is now commoditized to scientists that are not versed in AI model training.

Utilizing Synthetic Nodules for Improving Nodule Detection in Chest Radiographs.

Algorithms that automatically identify nodular patterns in chest X-ray (CXR) images could benefit radiologists by reducing reading time and improving accuracy. A promising approach is to use deep learning, where a deep neural network (DNN) is trained to classify and localize nodular patterns (including mass) in CXR images. Such algorithms, however, require enough abnormal cases to learn representations of nodular patterns arising in practical clinical settings. Obtaining large amounts of high-quality data is impractical in medical imaging where (1) acquiring labeled images is extremely expensive, (2) annotations are subject to inaccuracies due to the inherent difficulty in interpreting images, and (3) normal cases occur far more frequently than abnormal cases. In this work, we devise a framework to generate realistic nodules and demonstrate how they can be used to train a DNN identify and localize nodular patterns in CXR images. While most previous research applying generative models to medical imaging are limited to generating visually plausible abnormalities and using these patterns for augmentation, we go a step further to show how the training algorithm can be adjusted accordingly to maximally benefit from synthetic abnormal patterns. A high-precision detection model was first developed and tested on internal and external datasets, and the proposed method was shown to enhance the model's recall while retaining the low level of false positives.

Optimal number of strong labels for curriculum learning with convolutional neural network to classify pulmonary abnormalities in chest radiographs.

BACKGROUND AND OBJECTIVE: It is important to alleviate annotation efforts and costs by efficiently training on medical images. We performed a stress test on several strong labels for curriculum learning with a convolutional neural network to differentiate normal and five types of pulmonary abnormalities in chest radiograph images. METHODS: The numbers of CXR images of healthy subjects and patients, acquired at Asan Medical Center (AMC), were 6069 and 3465, respectively. The numbers of CXR images of patients with nodules, consolidation, interstitial opacity, pleural effusion, and pneumothorax were 944, 550, 280, 1360, and 331, respectively. The AMC dataset was split into training, tuning, and test, with a ratio of 7:1:2. All lesions were strongly labeled by thoracic expert radiologists, with confirmation of the corresponding CT. For curriculum learning, normal and abnormal patches (N = 26658) were randomly extracted around the normal lung and strongly labeled abnormal lesions, respectively. In addition, 1%, 5%, 20%, 50%, and 100% of strong labels were used to determine an optimal number for them. Each patch dataset was trained with the ResNet-50 architecture, and all CXRs with weak labels were used for fine-tuning them in a transfer-learning manner. A dataset acquired from the Seoul National University Bundang Hospital (SNUBH) was used for external validation. RESULTS: The detection accuracies of the 1%, 5%, 20%, 50%, and 100% datasets were 90.51, 92.15, 93.90, 94.54, and 95.39, respectively, in the AMC dataset and 90.01, 90.14, 90.97, 91.92, and 93.00 in the SNUBH dataset. CONCLUSIONS: Our results showed that curriculum learning with over 20% sampling rate for strong labels are sufficient to train a model with relatively high performance, which can be easily and efficiently developed in an actual clinical setting.

Electrical Characterization of the Tongue and the Soft Palate Using Lumped-Element Model for Intraoral Neuromodulation.

Intraoral functions are results of complex and well-orchestrated sensorimotor loop operations, and therefore vulnerable to small functional or neural defects. To secure the vital intraoral functions, it is important to find a way to favorably intervene the intraoral sensorimotor loop operations. The tongue and the soft palate are heavily associated with intraoral sensorimotor loops, with their dense neural innervations and occupancy of intraoral space. Therefore, electrical stimulation onto the tongue and the soft palate has a great potential to solve the problems in the intraoral functions. However, the electrical interface for both of them have not been characterized yet as a lumped-element model, for designing electrical stimulation and analyzing its effect. In this study, we measured stimulation thresholds to evoke electrotactile feedback and characterized electrical impedance across electrodes using lumped-element models. We found that average perception/discomfort thresholds for the tongue tip, lateral-inferior side of the tongue, and anterolateral side of the soft palate as 0.18/1.31, 0.37/3.99, and 1.19/7.55 mA, respectively. An R-C-R-R-C model represented the electrical interface across the tongue and the soft palate with the highest accuracy. The average component values of the R-C-R-R-C model were found as 2.72kΩ, 45.25nF, 1.27kΩ, 22.09GΩ, and 53.00nF, on average.

Added Value of Deep Learning-based Detection System for Multiple Major Findings on Chest Radiographs: A Randomized Crossover Study.

Background Previous studies assessing the effects of computer-aided detection on observer performance in the reading of chest radiographs used a sequential reading design that may have biased the results because of reading order or recall bias. Purpose To compare observer performance in detecting and localizing major abnormal findings including nodules, consolidation, interstitial opacity, pleural effusion, and pneumothorax on chest radiographs without versus with deep learning-based detection (DLD) system assistance in a randomized crossover design. Materials and Methods This study included retrospectively collected normal and abnormal chest radiographs between January 2016 and December 2017 (https://cris.nih.go.kr/; registration no. KCT0004147). The radiographs were randomized into two groups, and six observers, including thoracic radiologists, interpreted each radiograph without and with use of a commercially available DLD system by using a crossover design with a washout period. Jackknife alternative free-response receiver operating characteristic (JAFROC) figure of merit (FOM), area under the receiver operating characteristic curve (AUC), sensitivity, specificity, false-positive findings per image, and reading times of observers with and without the DLD system were compared by using McNemar and paired t tests. Results A total of 114 normal (mean patient age ± standard deviation, 51 years ± 11; 58 men) and 114 abnormal (mean patient age, 60 years ± 15; 75 men) chest radiographs were evaluated. The radiographs were randomized to two groups: group A (n = 114) and group B (n = 114). Use of the DLD system improved the observers' JAFROC FOM (from 0.90 to 0.95, P = .002), AUC (from 0.93 to 0.98, P = .002), per-lesion sensitivity (from 83% [822 of 990 lesions] to 89.1% [882 of 990 lesions], P = .009), per-image sensitivity (from 80% [548 of 684 radiographs] to 89% [608 of 684 radiographs], P = .009), and specificity (from 89.3% [611 of 684 radiographs] to 96.6% [661 of 684 radiographs], P = .01) and reduced the reading time (from 10-65 seconds to 6-27 seconds, P < .001). The DLD system alone outperformed the pooled observers (JAFROC FOM: 0.96 vs 0.90, respectively, P = .007; AUC: 0.98 vs 0.93, P = .003). Conclusion Observers including thoracic radiologists showed improved performance in the detection and localization of major abnormal findings on chest radiographs and reduced reading time with use of a deep learning-based detection system. © RSNA, 2021 Online supplemental material is available for this article.

Content-Based Image Retrieval of Chest CT with Convolutional Neural Network for Diffuse Interstitial Lung Disease: Performance Assessment in Three Major Idiopathic Interstitial Pneumonias.

OBJECTIVE: To assess the performance of content-based image retrieval (CBIR) of chest CT for diffuse interstitial lung disease (DILD). MATERIALS AND METHODS: The database was comprised by 246 pairs of chest CTs (initial and follow-up CTs within two years) from 246 patients with usual interstitial pneumonia (UIP, n = 100), nonspecific interstitial pneumonia (NSIP, n = 101), and cryptogenic organic pneumonia (COP, n = 45). Sixty cases (30-UIP, 20-NSIP, and 10-COP) were selected as the queries. The CBIR retrieved five similar CTs as a query from the database by comparing six image patterns (honeycombing, reticular opacity, emphysema, ground-glass opacity, consolidation and normal lung) of DILD, which were automatically quantified and classified by a convolutional neural network. We assessed the rates of retrieving the same pairs of query CTs, and the number of CTs with the same disease class as query CTs in top 1-5 retrievals. Chest radiologists evaluated the similarity between retrieved CTs and queries using a 5-scale grading system (5-almost identical; 4-same disease; 3-likelihood of same disease is half; 2-likely different; and 1-different disease). RESULTS: The rate of retrieving the same pairs of query CTs in top 1 retrieval was 61.7% (37/60) and in top 1-5 retrievals was 81.7% (49/60). The CBIR retrieved the same pairs of query CTs more in UIP compared to NSIP and COP (p = 0.008 and 0.002). On average, it retrieved 4.17 of five similar CTs from the same disease class. Radiologists rated 71.3% to 73.0% of the retrieved CTs with a similarity score of 4 or 5. CONCLUSION: The proposed CBIR system showed good performance for retrieving chest CTs showing similar patterns for DILD.

Development of machine learning-based clinical decision support system for hepatocellular carcinoma.

There is a significant discrepancy between the actual choice for initial treatment option for hepatocellular carcinoma (HCC) and recommendations from the currently used BCLC staging system. We develop a machine learning-based clinical decision support system (CDSS) for recommending initial treatment option in HCC and predicting overall survival (OS). From hospital records of 1,021 consecutive patients with HCC treated at a single centre in Korea between January 2010 and October 2010, we collected information on 61 pretreatment variables, initial treatment, and survival status. Twenty pretreatment key variables were finally selected. We developed the CDSS from the derivation set (N = 813) using random forest method and validated it in the validation set (N = 208). Among the 1,021 patients (mean age: 56.9 years), 81.8% were male and 77.0% had positive hepatitis B BCLC stages 0, A, B, C, and D were observed in 13.4%, 26.0%, 18.0%, 36.6%, and 6.3% of patients, respectively. The six multi-step classifier model was developed for treatment decision in a hierarchical manner, and showed good performance with 81.0% of accuracy for radiofrequency ablation (RFA) or resection versus not, 88.4% for RFA versus resection, and 76.8% for TACE or not. We also developed seven survival prediction models for each treatment option. Our newly developed HCC-CDSS model showed good performance in terms of treatment recommendation and OS prediction and may be used as a guidance in deciding the initial treatment option for HCC.

Prediction of Adverse Events in Stable Non-Variceal Gastrointestinal Bleeding Using Machine Learning.

Clinical risk-scoring systems are important for identifying patients with upper gastrointestinal bleeding (UGIB) who are at a high risk of hemodynamic instability. We developed an algorithm that predicts adverse events in patients with initially stable non-variceal UGIB using machine learning (ML). Using prospective observational registry, 1439 out of 3363 consecutive patients were enrolled. Primary outcomes included adverse events such as mortality, hypotension, and rebleeding within 7 days. Four machine learning algorithms, namely, logistic regression with regularization (LR), random forest classifier (RF), gradient boosting classifier (GB), and voting classifier (VC), were compared with the Glasgow-Blatchford score (GBS) and Rockall scores. The RF model showed the highest accuracies and significant improvement over conventional methods for predicting mortality (area under the curve: RF 0.917 vs. GBS 0.710), but the performance of the VC model was best in hypotension (VC 0.757 vs. GBS 0.668) and rebleeding within 7 days (VC 0.733 vs. GBS 0.694). Clinically significant variables including blood urea nitrogen, albumin, hemoglobin, platelet, prothrombin time, age, and lactate were identified by the global feature importance analysis. These results suggest that ML models will be useful early predictive tools for identifying high-risk patients with initially stable non-variceal UGIB admitted at an emergency department.

CT kernel conversions using convolutional neural net for super-resolution with simplified squeeze-and-excitation blocks and progressive learning among smooth and sharp kernels.

PURPOSE: Computed tomography (CT) volume sets reconstructed with different kernels are helping to increase diagnostic accuracy. However, several CT volumes reconstructed with different kernels are difficult to sustain, due to limited storage and maintenance issues. A CT kernel conversion method is proposed using convolutional neural networks (CNN). METHODS: A total of 3289 CT images from ten patients (five men and five women; mean age, 63.0 ± 8.6 years) were obtained in May 2016 (Somatom Sensation 16, Siemens Medical Systems, Forchheim, Germany). These CT images were reconstructed with various kernels, including B10f (very smooth), B30f (medium smooth), B50f (medium sharp), and B70f (very sharp) kernels. Smooth kernel images were converted into sharp kernel images using super-resolution (SR) network with Squeeze-and-Excitation (SE) blocks and auxiliary losses, and vice versa. In this study, the single-conversion model and multi-conversion model were presented. In case of the single-conversion model, for the one corresponding output image (e.g., B10f to B70), SE-Residual blocks were stacked. For the multi-conversion model, to convert an image into several output images (e.g., B10f to B30f, B50f, and B70f, and vice versa), progressive learning (PL) was employed by calculating auxiliary losses in every four SE-Residual blocks. Through auxiliary losses, the model could learn mutual relationships between different kernel types. The conversion quality was evaluated by the root-mean-square-error (RMSE), structural similarity (SSIM) index and mutual information (MI) between original and converted images. RESULTS: The RMSE (SSIM index , MI) of the multi-conversion model was 4.541 ± 0.688 (0.998 ± 0.001 , 2.587 ± 0.137), 27.555 ± 5.876 (0.944 ± 0.021 , 1.735 ± 0.137), 72.327 ± 17.387 (0.815 ± 0.053 , 1.176 ± 0.096), 8.748 ± 1.798 (0.996 ± 0.002 , 2.464 ± 0.121), 9.470 ± 1.772 (0.994 ± 0.003 , 2.336 ± 0.133), and 9.184 ± 1.605 (0.994 ± 0.002 , 2.342 ± 0.138) in conversion between B10f-B30f, B10f-B50f, B10f-B70f, B70f-B50f, B70f-B30f, and B70f-B10f, respectively, which showed significantly better image quality than the conventional model. CONCLUSIONS: We proposed deep learning-based CT kernel conversion using SR network. By introducing simplified SE blocks and PL, the model performance was significantly improved.

Active learning for accuracy enhancement of semantic segmentation with CNN-corrected label curations: Evaluation on kidney segmentation in abdominal CT.

Segmentation is fundamental to medical image analysis. Recent advances in fully convolutional networks has enabled automatic segmentation; however, high labeling efforts and difficulty in acquiring sufficient and high-quality training data is still a challenge. In this study, a cascaded 3D U-Net with active learning to increase training efficiency with exceedingly limited data and reduce labeling efforts is proposed. Abdominal computed tomography images of 50 kidneys were used for training. In stage I, 20 kidneys with renal cell carcinoma and four substructures were used for training by manually labelling ground truths. In stage II, 20 kidneys from the previous stage and 20 newly added kidneys were used with convolutional neural net (CNN)-corrected labelling for the newly added data. Similarly, in stage III, 50 kidneys were used. The Dice similarity coefficient was increased with the completion of each stage, and shows superior performance when compared with a recent segmentation network based on 3D U-Net. The labeling time for CNN-corrected segmentation was reduced by more than half compared to that in manual segmentation. Active learning was therefore concluded to be capable of reducing labeling efforts through CNN-corrected segmentation and increase training efficiency by iterative learning with limited data.

Endoscopic diagnosis and treatment planning for colorectal polyps using a deep-learning model.

We aimed to develop a computer-aided diagnostic system (CAD) for predicting colorectal polyp histology using deep-learning technology and to validate its performance. Near-focus narrow-band imaging (NBI) pictures of colorectal polyps were retrieved from the database of our institution. Of these, 12480 image patches of 624 polyps were used as a training set to develop the CAD. The CAD performance was validated with two test datasets of 545 polyps. Polyps were classified into three histological groups: serrated polyp (SP), benign adenoma (BA)/mucosal or superficial submucosal cancer (MSMC), and deep submucosal cancer (DSMC). The overall kappa value measuring the agreement between the true polyp histology and the expected histology by the CAD was 0.614-0.642, which was higher than that of trainees (n = 6, endoscopists with experience of 100 NBI colonoscopies in <6 months; 0.368-0.401) and almost comparable with that of the experts (n = 3, endoscopists with experience of 2,500 NBI colonoscopies in ≥5 years) (0.649-0.735). The areas under the receiver operating curves for CAD were 0.93-0.95, 0.86-0.89, and 0.89-0.91 for SP, BA/MSMC, and DSMC, respectively. The overall diagnostic accuracy of the CAD was 81.3-82.4%, which was significantly higher than that of the trainees (63.8-71.8%, P < 0.01) and comparable with that of experts (82.4-87.3%). The kappa value and diagnostic accuracies of the trainees improved with CAD assistance: that is, the kappa value increased from 0.368 to 0.655, and the overall diagnostic accuracy increased from 63.8-71.8% to 82.7-84.2%. CAD using a deep-learning model can accurately assess polyp histology and may facilitate the diagnosis of colorectal polyps by endoscopists.

Fully Automated Lung Lobe Segmentation in Volumetric Chest CT with 3D U-Net: Validation with Intra- and Extra-Datasets.

Lung lobe segmentation in chest CT has been used for the analysis of lung functions and surgical planning. However, accurate lobe segmentation is difficult as 80% of patients have incomplete and/or fake fissures. Furthermore, lung diseases such as chronic obstructive pulmonary disease (COPD) can increase the difficulty of differentiating the lobar fissures. Lobar fissures have similar intensities to those of the vessels and airway wall, which could lead to segmentation error in automated segmentation. In this study, a fully automated lung lobe segmentation method with 3D U-Net was developed and validated with internal and external datasets. The volumetric chest CT scans of 196 normal and mild-to-moderate COPD patients from three centers were obtained. Each scan was segmented using a conventional image processing method and manually corrected by an expert thoracic radiologist to create gold standards. The lobe regions in the CT images were then segmented using a 3D U-Net architecture with a deep convolutional neural network (CNN) using separate training, validation, and test datasets. In addition, 40 independent external CT images were used to evaluate the model. The segmentation results for both the conventional and deep learning methods were compared quantitatively to the gold standards using four accuracy metrics including the Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), mean surface distance (MSD), and Hausdorff surface distance (HSD). In internal validation, the segmentation method achieved high accuracy for the DSC, JSC, MSD, and HSD (0.97 ± 0.02, 0.94 ± 0.03, 0.69 ± 0.36, and 17.12 ± 11.07, respectively). In external validation, high accuracy was also obtained for the DSC, JSC, MSD, and HSD (0.96 ± 0.02, 0.92 ± 0.04, 1.31 ± 0.56, and 27.89 ± 7.50, respectively). This method took 6.49 ± 1.19 s and 8.61 ± 1.08 s for lobe segmentation of the left and right lungs, respectively. Although various automatic lung lobe segmentation methods have been developed, it is difficult to develop a robust segmentation method. However, the deep learning-based 3D U-Net method showed reasonable segmentation accuracy and computational time. In addition, this method could be adapted and applied to severe lung diseases in a clinical workflow.

A Curriculum Learning Strategy to Enhance the Accuracy of Classification of Various Lesions in Chest-PA X-ray Screening for Pulmonary Abnormalities.

We evaluated the efficacy of a curriculum learning strategy using two steps to detect pulmonary abnormalities including nodule[s], consolidation, interstitial opacity, pleural effusion, and pneumothorax with chest-PA X-ray (CXR) images from two centres. CXR images of 6069 healthy subjects and 3417 patients at AMC and 1035 healthy subjects and 4404 patients at SNUBH were obtained. Our approach involved two steps. First, the regional patterns of thoracic abnormalities were identified by initial learning of patch images around abnormal lesions. Second, Resnet-50 was fine-tuned using the entire images. The network was weakly trained and modified to detect various disease patterns. Finally, class activation maps (CAM) were extracted to localise and visualise the abnormal patterns. For average disease, the sensitivity, specificity, and area under the curve (AUC) were 85.4%, 99.8%, and 0.947, respectively, in the AMC dataset and 97.9%, 100.0%, and 0.983, respectively, in the SNUBH dataset. This curriculum learning and weak labelling with high-scale CXR images requires less preparation to train the system and could be easily extended to include various diseases in actual clinical environments. This algorithm performed well for the detection and classification of five disease patterns in CXR images and could be helpful in image interpretation.

Lung Segmentation on HRCT and Volumetric CT for Diffuse Interstitial Lung Disease Using Deep Convolutional Neural Networks.

A robust lung segmentation method using a deep convolutional neural network (CNN) was developed and evaluated on high-resolution computed tomography (HRCT) and volumetric CT of various types of diffuse interstitial lung disease (DILD). Chest CT images of 617 patients with various types of DILD, including cryptogenic organizing pneumonia (COP), usual interstitial pneumonia (UIP), and nonspecific interstitial pneumonia (NSIP), were scanned using HRCT (1-2-mm slices, 5-10-mm intervals) and volumetric CT (sub-millimeter thickness without intervals). Each scan was segmented using a conventional image processing method and then manually corrected by an expert thoracic radiologist to create gold standards. The lung regions in the HRCT images were then segmented using a two-dimensional U-Net architecture with the deep CNN, using separate training, validation, and test sets. In addition, 30 independent volumetric CT images of UIP patients were used to further evaluate the model. The segmentation results for both conventional and deep-learning methods were compared quantitatively with the gold standards using four accuracy metrics: the Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), mean surface distance (MSD), and Hausdorff surface distance (HSD). The mean and standard deviation values of those metrics for the HRCT images were 98.84 ± 0.55%, 97.79 ± 1.07%, 0.27 ± 0.18 mm, and 25.47 ± 13.63 mm, respectively. Our deep-learning method showed significantly better segmentation performance (p < 0.001), and its segmentation accuracies for volumetric CT were similar to those for HRCT. We have developed an accurate and robust U-Net-based DILD lung segmentation method that can be used for patients scanned with different clinical protocols, including HRCT and volumetric CT.

A Fully Automated System Using A Convolutional Neural Network to Predict Renal Allograft Rejection: Extra-validation with Giga-pixel Immunostained Slides.

Pathologic diagnoses mainly depend on visual scoring by pathologists, a process that can be time-consuming, laborious, and susceptible to inter- and/or intra-observer variations. This study proposes a novel method to enhance pathologic scoring of renal allograft rejection. A fully automated system using a convolutional neural network (CNN) was developed to identify regions of interest (ROIs) and to detect C4d positive and negative peritubular capillaries (PTCs) in giga-pixel immunostained slides. The performance of faster R-CNN was evaluated using optimal parameters of the novel method to enlarge the size of labeled masks. Fifty and forty pixels of the enlarged size images showed the best performance in detecting C4d positive and negative PTCs, respectively. Additionally, the feasibility of deep-learning-assisted labeling as independent dataset to enhance detection in this model was evaluated. Based on these two CNN methods, a fully automated system for renal allograft rejection was developed. This system was highly reliable, efficient, and effective, making it applicable to real clinical workflow.

A Perlin Noise-Based Augmentation Strategy for Deep Learning with Small Data Samples of HRCT Images.

Deep learning is now widely used as an efficient tool for medical image classification and segmentation. However, conventional machine learning techniques are still more accurate than deep learning when only a small dataset is available. In this study, we present a general data augmentation strategy using Perlin noise, applying it to pixel-by-pixel image classification and quantification of various kinds of image patterns of diffuse interstitial lung disease (DILD). Using retrospectively obtained high-resolution computed tomography (HRCT) images from 106 patients, 100 regions-of-interest (ROIs) for each of six classes of image patterns (normal, ground-glass opacity, reticular opacity, honeycombing, emphysema, and consolidation) were selected for deep learning classification by experienced thoracic radiologists. For extra-validation, the deep learning quantification of the six classification patterns was evaluated for 92 HRCT whole lung images for which hand-labeled segmentation masks created by two experienced radiologists were available. FusionNet, a convolutional neural network (CNN), was used for training, test, and extra-validation on classifications of DILD image patterns. The accuracy of FusionNet with data augmentation using Perlin noise (89.5%, 49.8%, and 55.0% for ROI-based classification and whole lung quantifications by two radiologists, respectively) was significantly higher than that with conventional data augmentation (82.1%, 45.7%, and 49.9%, respectively). This data augmentation strategy using Perlin noise could be widely applied to deep learning studies for image classification and segmentation, especially in cases with relatively small datasets.

Development of a Computer-Aided Differential Diagnosis System to Distinguish Between Usual Interstitial Pneumonia and Non-specific Interstitial Pneumonia Using Texture- and Shape-Based Hierarchical Classifiers on HRCT Images.

A computer-aided differential diagnosis (CADD) system that distinguishes between usual interstitial pneumonia (UIP) and non-specific interstitial pneumonia (NSIP) using high-resolution computed tomography (HRCT) images was developed, and its results compared against the decision of a radiologist. Six local interstitial lung disease patterns in the images were determined, and 900 typical regions of interest were marked by an experienced radiologist. A support vector machine classifier was used to train and label the regions of interest of the lung parenchyma based on the texture and shape characteristics. Based on the regional classifications of the entire lung using HRCT, the distributions and extents of the six regional patterns were characterized through their CADD features. The disease division index of every area fraction combination and the asymmetric index between the left and right lungs were also evaluated. A second SVM classifier was employed to classify the UIP and NSIP, and features were selected through sequential-forward floating feature selection. For the evaluation, 54 HRCT images of UIP (n = 26) and NSIP (n = 28) patients clinically diagnosed by a pulmonologist were included and evaluated. The classification accuracy was measured based on a fivefold cross-validation with 20 repetitions using random shuffling. For comparison, thoracic radiologists assessed each case using HRCT images without clinical information or diagnosis. The accuracies of the radiologists' decisions were 75 and 87%. The accuracies of the CADD system using different features ranged from 70 to 81%. Finally, the accuracy of the proposed CADD system after sequential-forward feature selection was 91%.