Equitable Skin Disease Prediction Using Transfer Learning and Domain Adaptation

TL;DR

This work rigorously evaluated the effectiveness of multiple pre-trained models using the Diverse Dermatology Images (DDI) dataset, which uniquely encompasses both underrepresented and common skin tones, making it an ideal benchmark for assessing the effectiveness of this approach.

Abstract

In the realm of dermatology, the complexity of diagnosing skin conditions manually necessitates the expertise of dermatologists. Accurate identification of various skin ailments, ranging from cancer to inflammatory diseases, is paramount. However, existing artificial intelligence (AI) models in dermatology face challenges, particularly in accurately diagnosing diseases across diverse skin tones, with a notable performance gap in darker skin. Additionally, the scarcity of publicly available, unbiased datasets hampers the development of inclusive AI diagnostic tools. To tackle the challenges in accurately predicting skin conditions across diverse skin tones, we employ a transfer-learning approach that capitalizes on the rich, transferable knowledge from various image domains. Our method integrates multiple pre-trained models from a wide range of sources, including general and specific medical images, to improve the robustness and inclusiveness of the skin condition predictions. We rigorously evaluated the effectiveness of these models using the Diverse Dermatology Images (DDI) dataset, which uniquely encompasses both underrepresented and common skin tones, making it an ideal benchmark for assessing our approach. Among all methods, Med-ViT emerged as the top performer due to its comprehensive feature representation learned from diverse image sources. To further enhance performance, we conducted domain adaptation using additional skin image datasets such as HAM10000. This adaptation significantly improved model performance across all models.

Figures (3)

• Figure 1: Model Architecture with Pre-training and Transfer Learning. The diagram illustrates the two-phase model development process. The upper section depicts pre-training on extensive datasets, enhancing foundational knowledge across varied domains. The lower section demonstrates fine-tuning on skin disease datasets, specifically adapting the model for binary classification of skin diseases. The RETFOUND model, exemplified here, undergoes enhancement through transfer learning by incorporating new layers. This approach culminates in a refined classification model adept at predicting skin diseases based on learned patterns and features.
• Figure 2: Training loss curve shown for 50 epochs (The training continues up to 100 epochs) during fine-tuning RETFound model on the DDI dataset.
• Figure 3: Comparative analysis of model performances across DDI and Ham10000+DDI datasets, highlighting the efficacy of domain adaptation and model scaling. The comparison includes a range of models: DeepDerm and Ham10000, specifically trained on skin disease images; MedViT, a model pretrained on general medical images; YOLOv8, adapted from general imaging tasks; and RETFOUND, originating from retinal image analyses. Notably, MedViT-base outperforms other models in adapting to both datasets, showcasing its robustness and versatility in domain adaptation scenarios.