JMIR Publications

ABSTRACT

Background:

While deep learning offers promising solutions for skin lesion classification, many models lack interpretability, limiting their clinical adoption. For example, distinguishing vitiligo from post-inflammatory hypopigmentation (PIH) can be clinically challenging due to their visual similarity.

Objective:

This study proposes an interpretable deep learning approach using a fine-tuned MobileNetV2 model to classify skin images of vitiligo and PIH.

Methods:

The model was trained and validated on a balanced dataset using five-fold cross-validation. To enhance clinical transparency, we applied an ensemble of visual explanation methods—Grad-CAM, Integrated Gradients, and SmoothGrad—providing complementary insights into model decisions and lesion localization.

Results:

Our results show that MobileNetV2 achieves robust performance across folds, effectively distinguishing between the two conditions. The combined use of three interpretability techniques highlighted key lesion features and revealed subtle pigmentation patterns, helping bridge the gap between automated prediction and clinical reasoning. Unlike prior models that prioritize accuracy at the expense of transparency, our approach strikes a balance between the two through a lightweight yet interpretable design.

Conclusions:

The ensemble of three explainability techniques not only strengthens clinical trust but also provides dermatologists with granular insights into depigmentation patterns, offering actionable visual cues. This makes our approach particularly valuable in early-stage diagnosis and in low-resource clinical environments where expert availability is limited.