NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Cascade Classification of Dermoscopic Images of Skin Neoplasms with Controllable Sensitivity and External Clinical Validation
TL;DR
A gap exists in the effective classification of skin neoplasms using deep learning, particularly in transferring models from open datasets to clinical settings. A cascade classification method was developed, utilizing architectures like ViT-B/16 and EfficientNetV2-S, to improve sensitivity and accuracy in diagnosing skin lesions.
✦ Why It Matters
Engineers can leverage cascade classification to enhance diagnostic accuracy in clinical applications of AI for skin cancer detection.
Key Takeaways
- Explore cascade classification methods for improved diagnostic accuracy.
- Validate AI models on clinical datasets before deployment.
- Consider architecture performance variations in clinical settings.
- Implement sensitivity control to align with clinical decision-making.
Full Summary
Deep learning has shown promise in classifying dermoscopic images of skin neoplasms, but challenges remain in generalizing models from open datasets to clinical environments. This study compared four architectures—Vision Transformer (ViT-B/16), Swin Transformer (Swin-S), ConvNeXt-S, and EfficientNetV2-S—across three classification schemes: binary (malignant/benign), single-stage four-class (benign, melanoma, squamous cell carcinoma, basal cell carcinoma), and a two-stage cascade approach.
The cascade method improved macro F1 scores by recovering malignant lesions misclassified as benign. Internal evaluations yielded ROC-AUC scores between 0.952 and 0.966, but performance on clinical datasets dropped significantly, indicating a generalization gap.
Notably, the ViT-B/16 architecture underperformed in binary classification on clinical data. These findings emphasize the importance of external validation and recalibration before deploying AI models in clinical practice.