Multi-Disease Classification of 13,667 Body CT Scans Using Weakly Supervised Deep Learning

Abstract

Background: Training deep learning classifiers typically requires massive amounts of manual annotation. Weak supervision may leverage existing medical data to classify multiple diseases and organ systems. Purpose: To design multi-disease classifiers for body computed tomography (CT) scans using automatically extracted labels from radiology text reports. Materials & Methods: This retrospective study deployed rule-based algorithms to extract 19,255 disease labels from reports of 13,667 body CT scans of 12,092 subjects for training. Using a 3D DenseVNet, three organ systems were segmented: lungs/pleura, liver/gallbladder, and kidneys/ureters. For each organ, a 3D convolutional neural network classified normality versus four common diseases. Testing was performed on an additional 2,158 CT volumes relative to 2,875 manually derived reference labels. Results: Manual validation of the extracted labels confirmed 91 to 99% accuracy. Performance using the receiver operating characteristic area under the curve (AUC) for lungs/pleura labels were as follows: atelectasis 0.77 (95% CI: 0.74 to 0.81), nodule 0.65 (0.61 to 0.69), emphysema 0.89 (0.86 to 0.92), effusion 0.97 (0.96 to 0.98), and normal 0.89 (0.87 to 0.91). For liver/gallbladder: stone 0.62 (0.56 to 0.67), lesion 0.73 (0.69 to 0.77), dilation 0.87 (0.84 to 0.90), fatty 0.89 (0.86 to 0.92), and normal 0.82 (0.78 to 0.85). For kidneys/ureters: stone 0.83 (0.79 to 0.87), atrophy 0.92 (0.89 to 0.94), lesion 0.68 (0.64 to 0.72), cyst 0.70 (0.66 to 0.73), and normal 0.79 (0.75 to 0.83). Conclusion: Weakly supervised deep learning classifiers leveraged massive amounts of unannotated body CT data to classify multiple organ systems and diverse diseases.

Ähnliche Arbeiten