AI Model Differentiates BCC vs cSCC Subtypes
External validation identifies calibration shift in COBRA cohort.
-
By
-
Kathryn Wighton
-
March 23, 2026
-
Clinical Scorecard: AI Model Differentiates BCC vs cSCC Subtypes
At a Glance
| Category | Detail |
|---|
| Condition | Basal Cell Carcinoma (BCC) and Cutaneous Squamous Cell Carcinoma (cSCC) |
| Key Mechanisms | Weakly supervised deep learning model using clustering-constrained attention and feature extraction from a vision transformer. |
| Target Population | Patients with infiltrative basal cell carcinoma and poorly differentiated cutaneous squamous cell carcinoma. |
| Care Setting | Dermatopathology |
Key Highlights
- Model achieved 100% accuracy, sensitivity, and specificity on internal test set.
- External validation showed AUC of 1.0 in Queensland cohort and 0.92 in COBRA cohort.
- Attention heatmaps indicated tumor localization in histopathology images.
- Fine-tuning of HistoGPT model improved accuracy to 98% with high sensitivity and specificity.
- Calibration and domain adaptation are crucial for reliable deployment across institutions.
Guideline-Based Recommendations
Diagnosis
- Utilize weakly supervised deep learning models for accurate classification of BCC and cSCC.
Management
- Implement careful calibration and domain adaptation for model deployment.
Monitoring & Follow-up
- Regularly assess model performance across different cohorts and settings.
Risks
- Potential for calibration shifts and performance variation due to diagnostic subtype differences and image quality.
Patient & Prescribing Data
Patients with diagnosed BCC and cSCC requiring histopathological evaluation.
Deep learning models can enhance diagnostic accuracy in challenging cases.
Clinical Best Practices
- Ensure thorough validation of AI models in diverse clinical settings.
- Monitor model performance and adjust thresholds based on specific cohort characteristics.
References
