Automated Skin Cancer Report Generation via a Knowledge-Distilled Vision-Language Model

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Artificial Intelligence (AI) shows promise in analyzing dermoscopic images for skin cancer diagnostics.
• Lack of transparency and interpretability in AI diagnostic tools hinders clinical adoption.
• Automated report generation can improve AI explainability and reduce medical professional workload.

Purpose of the Study:

• To develop a multimodal vision-language model (VLM) for generating structured medical reports from dermoscopic images.
• To enhance the interpretability and clinical utility of AI-driven dermatological diagnostics.
• To bridge the gap between AI capabilities and clinical needs through automated report generation.

Main Methods:

• A two-stage knowledge distillation (KD) framework was employed to train the VLM.
• The model generates reports structured into Findings, Impression, and Differential Diagnosis sections.
• Reports incorporate descriptive features based on the 7-point melanoma checklist.

Main Results:

• The VLM successfully produced accurate and interpretable dermatological reports.
• Human feedback confirmed the clinical relevance, completeness, and interpretability of the generated reports.
• Computational metrics (SacreBLEU, ROUGE-1, ROUGE-L, BERTScore F1) validated report accuracy and alignment.

Conclusions:

• The multimodal VLM effectively generates structured, clinically relevant reports from dermoscopic images.
• The system's explainability and generalization capabilities are supported by its design and validation.
• This approach offers a significant advancement in AI-assisted skin cancer diagnostics.