Deep Neural Network for Early Image Diagnosis of Stevens-Johnson Syndrome/Toxic Epidermal Necrolysis

Background Stevens-Johnson syndrome (SJS)/toxic epidermal necrolysis (TEN) is a life-threatening cutaneous adverse drug reaction (cADR). Distinguishing SJS/TEN from nonsevere cADRs is difficult, especially in the early stages of the disease. Objective To overcome this limitation, we developed a computer-aided diagnosis system for the early diagnosis of SJS/TEN, powered by a deep convolutional neural network (DCNN). Methods We trained a DCNN using a dataset of 26,661 individual lesion images obtained from 123 patients with a diagnosis of SJS/TEN or nonsevere cADRs. The DCNN's accuracy of classification was compared with that of 10 board-certified dermatologists and 24 trainee dermatologists. Results The DCNN achieved 84.6% sensitivity (95% confidence interval [CI], 80.6-88.6), whereas the sensitivities of the board-certified dermatologists and trainee dermatologists were 31.3 % (95% CI, 20.9-41.8; P < .0001) and 27.8% (95% CI, 22.6-32.5; P < .0001), respectively. The negative predictive value was 94.6% (95% CI, 93.2-96.0) for the DCNN, 68.1% (95% CI, 66.1-70.0; P < .0001) for the board-certified dermatologists, and 67.4% (95% CI, 66.1-68.7; P < .0001) for the trainee dermatologists. The area under the receiver operating characteristic curve of the DCNN for a SJS/TEN diagnosis was 0.873, which was significantly higher than that for all board-certified dermatologists and trainee dermatologists. Conclusions We developed a DCNN to classify SJS/TEN and nonsevere cADRs based on individual lesion images of erythema. The DCNN performed significantly better than did dermatologists in classifying SJS/TEN from skin images. Stevens-Johnson syndrome (SJS)/toxic epidermal necrolysis (TEN) is a life-threatening cutaneous adverse drug reaction (cADR). Distinguishing SJS/TEN from nonsevere cADRs is difficult, especially in the early stages of the disease. To overcome this limitation, we developed a computer-aided diagnosis system for the early diagnosis of SJS/TEN, powered by a deep convolutional neural network (DCNN). We trained a DCNN using a dataset of 26,661 individual lesion images obtained from 123 patients with a diagnosis of SJS/TEN or nonsevere cADRs. The DCNN's accuracy of classification was compared with that of 10 board-certified dermatologists and 24 trainee dermatologists. The DCNN achieved 84.6% sensitivity (95% confidence interval [CI], 80.6-88.6), whereas the sensitivities of the board-certified dermatologists and trainee dermatologists were 31.3 % (95% CI, 20.9-41.8; P < .0001) and 27.8% (95% CI, 22.6-32.5; P < .0001), respectively. The negative predictive value was 94.6% (95% CI, 93.2-96.0) for the DCNN, 68.1% (95% CI, 66.1-70.0; P < .0001) for the board-certified dermatologists, and 67.4% (95% CI, 66.1-68.7; P < .0001) for the trainee dermatologists. The area under the receiver operating characteristic curve of the DCNN for a SJS/TEN diagnosis was 0.873, which was significantly higher than that for all board-certified dermatologists and trainee dermatologists. We developed a DCNN to classify SJS/TEN and nonsevere cADRs based on individual lesion images of erythema. The DCNN performed significantly better than did dermatologists in classifying SJS/TEN from skin images.