Interpreting Biomedical VLMs on High-Imbalance Out-of-Distributions: An Insight into BiomedCLIP on Radiology

2025/06/17 by Nafiz Sadman, Sadman, Nafiz, Farhana Zulkernine +3
Medicine · #Computer Vision and Pattern Recognition (cs.CV) #ECG Monitoring and Analysis #FOS: Computer and information sciences #Lung Cancer Diagnosis and Treatment

paper · pdf · doi:10.48550/arxiv.2506.14136

openalex publication_date 2025/06/17 · openalex created_date 2025/10/18 · openalex updated_date 2026/07/28

Abstract

In this paper, we construct two research objectives: i) explore the learned embedding space of BiomedCLIP, an open-source large vision language model, to analyse meaningful class separations, and ii) quantify the limitations of BiomedCLIP when applied to a highly imbalanced, out-of-distribution multi-label medical dataset. We experiment on IU-xray dataset, which exhibits the aforementioned criteria, and evaluate BiomedCLIP in classifying images (radiographs) in three contexts: zero-shot inference, full finetuning, and linear probing. The results show that the model under zero-shot settings over-predicts all labels, leading to poor precision and inter-class separability. Full fine-tuning improves classification of distinct diseases, while linear probing detects overlapping features. We demonstrate visual understanding of the model using Grad-CAM heatmaps and compare with 15 annotations by a radiologist. We highlight the need for careful adaptations of the models to foster reliability and applicability in a real-world setting. The code for the experiments in this work is available and maintained on GitHub.

Citations

• A Multimodal Biomedical Foundation Model Trained from Fifteen Million Image–Text Pairs
• BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models
• Overcoming Data Scarcity in Biomedical Imaging with a Foundational Multi-Task Model
• Knowledge-enhanced Visual-Language Pre-training on Chest Radiology Images
• Finetune like you pretrain: Improved finetuning of zero-shot vision models
• Medical Image Segmentation Review: The success of U-Net
• Medical Image Segmentation Review: The Success of U-Net
• MedCLIP: Contrastive Learning from Unpaired Medical Images and Text
• Learning Transferable Visual Models From Natural Language Supervision
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Contrastive Learning of Medical Visual Representations from Paired Images and Text
• Domain-Specific Language Model Pretraining for Biomedical Natural Language Processing
• Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
• Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning
• A Review on Multi-Label Learning Algorithms

Related

• BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs
• Delving into Out-of-Distribution Detection with Medical Vision-Language Models
• Few-Shot, Now for Real: Medical VLMs Adaptation without Balanced Sets or Validation
• UniMed-CLIP: Towards a Unified Image-Text Pretraining Paradigm for\n Diverse Medical Imaging Modalities
• Anatomy-VLM: A Fine-grained Vision-Language Model for Medical Interpretation