The art of diagnosing rare skin tumors: Can DL-CNNs enhance dermatologists' diagnostic accuracy?

Importance: Deep learning convolutional neural networks (DL-CNN) achieved diagnostic accuracies comparable to dermatologists in controlled test environments. However, their performance in diagnosing rare skin tumors (RST) remains unclear. This study aimed to evaluate a binary DL-CNN's diagnostic performance in RST and assess the level of support for an international group of dermatologists.

Methods: In a cross-sectional reader study, a market-approved binary DL-CNN (Moleanalyzer-Pro) assessed 200 dermoscopic images in a conveniance sample of histologically confirmed RST. An international panel of dermatologists rated malignancy and management across three levels: (I) dermoscopy only, (II) dermoscopy, close-up images, and metadata, (III) level-II plus DL-CNN malignancy predictions. Sensitivity, specificity, and the area under the receiver operating characteristic curve (ROC-AUC) for the DL-CNN versus dermatologists (level-II). Secondary outcomes included performance changes across study levels.

Results: The DL-CNN achieved a sensitivity (95 % CI) of 66.7 % (56.4 %-75.6 %), specificity of 56.4 % (47.0 %-65.3 %), and ROC-AUC of 0.634 (0.557-0.711). Dermatologists reached a significantly higher mean sensitivity (80.3 %, 77.3 %-83.4 %), specificity (65.1 %, 61.3 %-69.0 %), and ROC-AUC (0.839, 0.783-0.894; all p < 0.001). With DL-CNN predictions, dermatologists' sensitivity slightly increased (81.3 %, p = 0.032), specificity decreased (64.0 %, p = 0.036), and ROC-AUC remained unchanged. The DL-CNN could not improve dermatologists' accuracy in misclassified cases.

Conclusion: The tested DL-CNN showed a limited diagnostic performance in diagnosing RST. While minor effects on expert decision-making were observed, overall diagnostic accuracy remained highest with full clinical context. Better training data are needed for improved DL-CNN performance in RST.