Deep learning in assisting dermatologists in classifying basal cell carcinoma from seborrheic keratosis

Li-Hong Mei; Meng-Ke Cao; Jing Li; Xuan-Guang Ye; Xiang-Dong Liu; Gao Yang

doi:10.3389/fonc.2025.1507322

Deep learning in assisting dermatologists in classifying basal cell carcinoma from seborrheic keratosis

Front Oncol. 2025 Apr 24:15:1507322. doi: 10.3389/fonc.2025.1507322. eCollection 2025.

Authors

Li-Hong Mei^#¹, Meng-Ke Cao^#², Jing Li^#², Xuan-Guang Ye³, Xiang-Dong Liu⁴, Gao Yang¹

Affiliations

¹ Department of Dermatology, Jinshan Hospital of Fudan University, Shanghai, China.
² Department of Dermatology, Shanghai Eighth People's Hospital, Shanghai, China.
³ Department of Pathology, Jinshan Hospital, Fudan University, Shanghai, China.
⁴ Department of Dermatology, Shanghai Sixth People's Hospital, Jiaotong University, Shanghai, China.

^# Contributed equally.

Abstract

Objectives: This study aimed to evaluate the effectiveness of deep learning model in assisting dermatologists in classifying basal cell carcinoma (BCC) from seborrheic keratosis (SK). The goal was to assess whether AI-assisted diagnostics could improve accuracy, reduce misdiagnoses, and potentially enhance clinical outcomes.

Methods: This prospective study included 707 patients with histopathologically confirmed BCC or SK as an internal dataset (validation cohort), along with 5572 patients from the ISIC public dataset as an external dataset (split into training and test cohort). The images were preprocessed and augmented before being fed into a deep learning model based on the CLIP ViT-B/16 architecture. The model's performance was assessed using the area under the receiver operating characteristic (ROC) curves (AUC). Two dermatologists, one with 3 years of experience and another with 15 years of experience, reviewed the cases before and after receiving the deep learning model's predictions. Net reclassification index (NRI) and integrated discrimination improvement (IDI), was used to quantify the improvement in reclassification performance.

Results: The model achieved an AUC of 0.76 in the training cohort and 0.79 in the test cohort for differentiating between BCC and SK. In the validation cohort, the model demonstrated an AUC of 0.71. Dermatologist 1's AUC improved from 0.75 to 0.82 with deep learning model assistance, while Dermatologist 2's AUC increased from 0.79 to 0.82. NRI and IDI analysis revealed statistically significant improvements, with Dermatologist 1 showing a 18% improvement and Dermatologist 2 showing a 11% improvement. Additionally, attention mechanisms like Grad-CAM provided insights into the model's decision-making process, enhancing the interpretability of its predictions.

Conclusion: The deep learning model demonstrated significant potential in aiding dermatologists in classifying BCC from SK.

Keywords: basal cell carcinoma; deep learning; dermatologist assistance; seborrheic keratosis; skin cancer diagnosis.