Comparative Development of BioMedCLIP for Enhanced Biomedical Data Integration

Advancements in multimodal learning in Artificial Intelligence (AI) have led to the proposal of many foundational models in the area of biomedical AI. However, leveraging these models for specialized clinical use requires a satisfactory level of fine-tuning and validation. BioMedCLIP is one such state-of-the-art model. This paper presents a comparative study that focuses on adapting BioMedCLIP and evaluating its efficiency across various datasets and fine-tuning approaches. Several fine-tuning methods were explored, and the model was implemented on two large and challenging datasets. The first dataset was the large-scale National Institutes of Health (NIH) Chest X-ray collection for multi-label disease classification; the second dataset was HAM10000, a large collection of multi-source dermatoscopic images of skin lesions. The primary objective was to assess different fine-tuning strategies and develop a specialized model with enhanced capabilities in integrating images and textual data. Using three different approaches, the study compares BioMedCLIP across the NIH Chest X-ray and HAM10000 datasets, demonstrating improved text–image modality integration, 55.6% macro accuracy, and reduced overfitting. This work validates the effectiveness of domain-specific adaptation and establishes a powerful, fine-tuned model ready for deployment in advanced healthcare applications.