Deep learning based histological classification of adnex tumors

Area of Science:

• Dermatopathology
• Computational Pathology
• Medical Artificial Intelligence

Background:

• Cutaneous adnexal tumors originate from hair appendages and range from benign to life-threatening malignant forms.
• Accurate diagnosis is crucial for effective patient treatment and outcomes.
• Artificial intelligence (AI), particularly deep neural networks, shows significant potential in medical diagnostics, including pathology for common skin tumors.

Purpose of the Study:

• To evaluate the capability of AI to accurately identify and differentiate a diverse range of cutaneous adnexal tumors.
• To assess AI's performance on less common tumor entities, expanding its diagnostic utility beyond frequent skin neoplasms.

Main Methods:

• A diverse set of cutaneous adnexal tumors was curated from five German pathology centers.
• A deep neural network algorithm was trained using samples from four centers and validated on data from a fifth independent center.

Main Results:

• The AI model successfully differentiated 14 distinct cutaneous adnexal tumor types and common skin tumors like basal cell carcinoma and seborrheic keratosis.
• The algorithm achieved a total accuracy of 89.92% in classifying 248 samples into 16 diagnostic categories.
• AI demonstrated proficiency in distinguishing rare tumor entities, even when trained with limited case numbers (<50).

Conclusions:

• The study highlights the substantial potential of AI as a tool to assist pathologists in routine dermatopathology diagnostics.
• AI can reliably distinguish between various cutaneous adnexal tumors, including rare entities.
• While AI shows promise, the ultimate diagnostic responsibility remains with the human pathologist.