Detection of Pneumothorax with Deep Learning Models: Learning From Radiologist Labels vs Natural Language Processing Model Generated Labels

Rationale and Objectives: To compare the performance of pneumothorax deep learning detection models trained with radiologist versus natural language processing (NLP) labels on the NIH ChestX-ray14 dataset. Materials and Methods: The ChestX-ray14 dataset consisted of 112,120 frontal chest radiographs with 5302 positive and 106, 818 negative labels for pneumothorax using NLP (dataset A). All 112,120 radiographs were also inspected by 4 radiologists leaving a visually confirmed set of 5,138 positive and 104,751 negative for pneumothorax (dataset B). Datasets A and B were used independently to train 3 convolutional neural network (CNN) architectures (ResNet-50, DenseNet-121 and EfficientNetB3). All models' area under the receiver operating characteristic curve (AUC) were evaluated with the official NIH test set and an external test set of 525 chest radiographs from our emergency department. Results: There were significantly higher AUCs on the NIH internal test set for CNN models trained with radiologist vs NLP labels across all architectures. AUCs for the NLP/radiologist-label models were 0.838 (95%CI:0.830, 0.846)/0.881 (95%CI:0.873,0.887) for ResNet-50 (p = 0.034), 0.839 (95%CI:0.831,0.847)/0.880 (95%CI:0.873,0.887) for DenseNet-121, and 0.869 (95%CI: 0.863,0.876)/0.943 (95%CI: 0.939,0.946) for EfficientNetB3 (p <= 0.001). Evaluation with the external test set also showed higher AUCs (p <0.001) for the CNN models trained with radiologist versus NLP labels across all architectures. The AUCs for the NLP/radiologist-label models were 0.686 (95%CI:0.632,0.740)/0.806 (95%CI:0.758,0.854) for ResNet-50, 0.736 (95%CI:0.686, 0.787)/0.871 (95%CI:0.830,0.912) for DenseNet-121, and 0.822 (95%CI: 0.775,0.868)/0.915 (95%CI: 0.882,0.948) for EfficientNetB3. Conclusion: We demonstrated improved performance and generalizability of pneumothorax detection deep learning models trained with radiologist labels compared to models trained with NLP labels.