AI Model Differentiates BCC vs cSCC Subtypes

External validation identifies calibration shift in COBRA cohort.

By
Kathryn Wighton
March 23, 2026
3 min

Conexiant

Overview

Replace 'perfect classification' with 'highly accurate classification' to reflect clinical realities.

Background

Accurate histopathologic subtyping is crucial for risk stratification and treatment selection in nonmelanoma skin cancers. Differentiating between BCC and cSCC subtypes can significantly influence management strategies, including surgical and nonsurgical options. Recent advancements in artificial intelligence (AI) offer promising tools for enhancing diagnostic accuracy in challenging cases.

Data Highlights

Dataset	AUC	Accuracy	Sensitivity	Specificity
Internal Test Set	1.0	100%	100%	100%
Queensland Cohort	1.0	90%	100%	90%
COBRA Cohort	0.92	52% (87% after adjustment)	100% (79% after adjustment)	4% (94% after adjustment)

Key Findings

The AI model achieved an AUC of 1.0 on the internal test set with 100% accuracy.
In the Queensland cohort, the model maintained an AUC of 1.0 with 90% accuracy.
In the COBRA cohort, accuracy improved from 52% to 87% after threshold adjustment using Youden’s J statistic.
Attention heatmaps indicated tumor localization, with distinct patterns for BCC and cSCC.
Fine-tuning the HistoGPT model on the in-house dataset resulted in an AUC of 1.0 with 98% accuracy.

Clinical Implications

The findings suggest that AI models can significantly enhance the accuracy of diagnosing BCC and cSCC, potentially leading to improved patient management. Careful calibration and domain adaptation are essential for reliable deployment of these models across different clinical settings.

Conclusion

This study highlights the potential of weakly supervised deep learning in accurately classifying challenging skin cancer subtypes, paving the way for enhanced diagnostic tools in dermatopathology.