Evaluation of the Performance of Deep Learning Models Trained on a Combination of Major Abnormal Patterns on Chest Radiographs for Major Chest Diseases at International Multi-Centers
To evaluate the abnormal classification performance for major chest diseases using a deep learning model that was trained on a combination of major abnormal patterns on chest radiographs.
METHOD AND MATERIALS
We experimented with the abnormal classification performance for a deep learning model for major diseases (tuberculosis and pneumonia) that was trained on a combination of different patterns (nodule, consolidation and interstitial opacity) on CRs. To evaluate the effect of each pattern combination on performance for major diseases, we tested five cases of patterns, which is composed of the nodule case, the consolidation case, the interstitial opacity case, the combination of consolidation and interstitial opacity case, and the combination of all three cases. When training each case, all normal data was used for training. CRs with three abnormal patterns and normal patterns were used as training datasets, which were received from two hospitals and consisted of 2095, 2401, 1290, and 3000 images for nodule, consolidation, interstitial opacity, and normal patterns, respectively. And all abnormal CRs were clinically confirmed by CT scans. For an explicit evaluation, the public dataset was used as the test dataset, which consists of the Shenzhen (normal: 326, tuberculosis: 336) and PadChest (normal: 300, pneumonia: 127, randomly selected) dataset, which was used to evaluate tuberculosis and pneumonia, respectively.
In the test dataset, for tuberculosis and pneumonia, the classification performance of the models trained with the five cases of patterns showed AUC 0.58 / 0.69 for nodule case, 0.76 / 0.82 for consolidation, 0.52 / 0.76 for interstitial opacity case, 0.79 / 0.83 for combination of consolidation and interstitial opacity case, 0.79 / 0.82 for combination of all three case, respectively.
We have shown through experimentations that the deep learning model trained from data with major patterns (nodule, consolidation, interstitial opacity) can classify major diseases (tuberculosis, pneumonia) as abnormal. Also, consolidation was highly correlated with tuberculosis and pneumonia. On the other hand, interstitial opacity and nodule were more correlated with pneumonia, tuberculosis, respectively.
The diagnosis based on the patterns of abnormal findings allows detection of various diseases.
Deep Learning-Based Automated Segmentation of Prostate Cancer on Multiparametric MRI: Comparison with Experienced Uroradiologists
To compare the performance of deep learning based prostate cancer (PCa) segmentation with manual segmentation of experienced uroradiologists.
METHOD AND MATERIALS
From 2011 Jan to 2018 Apr, 350 patients who underwent prostatectomy for prostate cancer were enrolled retrospectively. To collect histopathological ground truth, pathologic slides of whole resected prostate were scanned and PCa lesions were drawn by a uropathologist with 25 years' experience. With reference to the histopathological lesion, radiological ground truth of PCa was drawn on the T2 weighted image by a uroradiologist with 19 years' experience. A U-Net type deep neural network, in which the encoder part has more convolution blocks than the decoder, was trained for segmentation. Four different MR sequences including T2 weighted images, diffusion weighted images (b = 0, 1000), and apparent diffusion coefficient (ADC) images, were used as input images after affine registration. Besides the automatic segmentation by the deep neural network, two experienced uroradiologists marked suspected sectors of PCa among 39 sectors provided by PIRADS-v2 after reviewing same images of four MR sequences. The manual segmentation performance of uroradiologists was measured using the number of sectors that coincided with the ground truth PCa lesion.
The dice coefficient scores (DCSs) achieved by two uroradiologists were 0.490 and 0.310 respectively. The DCS was calculated based on the number of sectors. The DCS of automatic segmentation by a deep neural network was 0.558 (calculated by the number of pixels) which is slightly better than the average (0.40) DCSs of uroradiologists.
Automated segmentation of PCa on multiparametric MR based on histopathologically confirmed lesion label achieved comparable performance with experienced uroradiologist.
The automated segmentation of prostate cancer using a deep neural network not only reduce time consuming work but also provide reliable location and size information required for treatment decision.
Deep Learning Algorithm for Reducing CT Slice Thickness: Effect on Reproducibility of Radiomics in Lung Cancer
To retrospectively assess the effect of CT slice thickness on the reproducibility of radiomic features (RFs) of lung cancer, and to investigate if convolutional neural network (CNN)-based super-resolution (SR) algorithms can improve the reproducibility of RFs obtained from different slice thicknesses.
METHOD AND MATERIALS
CT images from 100 pathologically proven lung cancers acquired between July 2017 and December 2017 were evaluated, including 1, 3, and 5 mm slice thicknesses. CNN-based SR algorithms using residual learning were developed to convert thick-slice images into 1 mm slices. Lung cancers were semi-automatically segmented and a total of 702 RFs (tumor intensity, texture, and wavelet features) were extracted from 1, 3, and 5 mm slices, as well as the 1 mm slices generated from the 3 and 5 mm images. The stabilities of the RFs were evaluated using concordance correlation coefficients (CCCs).
All CT scans were successfully converted to 1 mm slice images at a rate of 2.5 s/slice. The mean CCCs for the comparisons of original 1 vs 3 mm, 1 vs 5 mm, and 3 vs 5 mm images were 0.41, 0.27, and 0.65, respectively (all, P<0.001). Tumor intensity features showed the best reproducibility and wavelets the lowest. The majority of RFs failed to reach reproducibility (CCC>=0.85; 3.6%, 1.0%, and 21.5%, respectively). In terms of nodule type, GGNs had better reproducibility than solid nodules in all RF classes and in all slice-thickness pairings (P < 0.001 for 1 vs 3 mm and 1 vs 5 mm, and P = 0.002 for 3 vs 5 mm). After applying CNN-based SR algorithms, the reproducibility significantly improved in all three pairings (mean CCCs: 0.58, 0.45, and 0.72; all, P<0.001). This improvement was also observed in the subgroupings based on the classes of RFs and nodule types. The reproducible RFs also increased (36.3%, 17.4%, and 36.9%, respectively).
The reproducibility of RFs in lung cancer is significantly influenced by CT slice thickness, which can be improved by the CNN-based SR algorithms.
On the basis of the findings of our study, the comparisons of radiomics results derived from CT images with different slice thicknesses may be unreliable. As our convolutional neural network-based image conversion algorithm is easily applicable and reliable, this algorithm may be used for enhancing reproducibility of radiomic features when the CT slice-thicknesses are different.
DeNTNet: Deep Neural Transfer Network for the detection of periodontal bone loss using panoramic dental radiographs
AbstractIn this study, a deep learning-based method for developing an automated diagnostic support system that detects periodontal bone loss in the panoramic dental radiographs is proposed. The presented method called DeNTNet not only detects lesions but also provides the corresponding teeth numbers of the lesion according to dental federation notation. DeNTNet applies deep convolutional neural networks(CNNs) using transfer learning and clinical prior knowledge to overcome the morphological variation of the lesions and imbalanced training dataset. With 12,179 panoramic dental radiographs annotated by experienced dental clinicians, DeNTNet was trained, validated, and tested using 11,189, 190, and 800 panoramic dental radiographs, respectively. Each experimental model was subjected to comparative study to demonstrate the validity of each phase of the proposed method. When compared to the dental clinicians, DeNTNet achieved the F1 score of 0.75 on the test set, whereas the average performance of dental clinicians was 0.69.
Deep Learning-based Detection System for Multiclass Lesions on Chest Radiographs: Comparison with Observer Readings
To investigate the feasibility of a deep learning–based detection (DLD) system for multiclass lesions on chest radiograph, in comparison with observers.
A total of 15,809 chest radiographs were collected from two tertiary hospitals (7204 normal and 8605 abnormal with nodule/mass, interstitial opacity, pleural effusion, or pneumothorax). Except for the test set (100 normal and 100 abnormal (nodule/mass, 70; interstitial opacity, 10; pleural effusion, 10; pneumothorax, 10)), radiographs were used to develop a DLD system for detecting multiclass lesions. The diagnostic performance of the developed model and that of nine observers with varying experiences were evaluated and compared using area under the receiver operating characteristic curve (AUROC), on a per-image basis, and jackknife alternative free-response receiver operating characteristic figure of merit (FOM) on a per-lesion basis. The false-positive fraction was also calculated.
Compared with the group-averaged observations, the DLD system demonstrated significantly higher performances on image-wise normal/abnormal classification and lesion-wise detection with pattern classification (AUROC, 0.985 vs. 0.958; p = 0.001; FOM, 0.962 vs. 0.886; p < 0.001). In lesion-wise detection, the DLD system outperformed all nine observers. In the subgroup analysis, the DLD system exhibited consistently better performance for both nodule/mass (FOM, 0.913 vs. 0.847; p < 0.001) and the other three abnormal classes (FOM, 0.995 vs. 0.843; p < 0.001). The false-positive fraction of all abnormalities was 0.11 for the DLD system and 0.19 for the observers.
The DLD system showed the potential for detection of lesions and pattern classification on chest radiographs, performing normal/abnormal classifications and achieving high diagnostic performance.