[Paper Review] GSCo: Towards Generalizable AI in Medicine via Generalist-Specialist Collaboration

Sunan He, Yuxiang Nie+22 more|arXiv (Cornell University)|Apr 23, 2024

Biomedical Text Mining and OntologiesBiochemistry, Genetics and Molecular Biology3 citations

TL;DR

This paper proposes GSCo, a generalist-specialist collaboration framework for medical AI that uses diagnosis-guided bootstrapping to generate high-quality vision-language data from medical images and labels. It introduces MedDr, a generalist foundation model trained on diverse modalities (radiology, pathology, dermatology, etc.), and enhances its generalization via a retrieval-augmented inference strategy, achieving state-of-the-art performance on medical VQA, report generation, and image diagnosis tasks, especially on rare diseases.

ABSTRACT

Generalist foundation models (GFMs) are renowned for their exceptional capability and flexibility in effectively generalizing across diverse tasks and modalities. In the field of medicine, while GFMs exhibit superior generalizability based on their extensive intrinsic knowledge as well as proficiency in instruction following and in-context learning, specialist models excel in precision due to their domain knowledge. In this work, for the first time, we explore the synergy between the GFM and specialist models, to enable precise medical image analysis on a broader scope. Specifically, we propose a cooperative framework, Generalist-Specialist Collaboration (GSCo), which consists of two stages, namely the construction of GFM and specialists, and collaborative inference on downstream tasks. In the construction stage, we develop MedDr, the largest open-source GFM tailored for medicine, showcasing exceptional instruction-following and in-context learning capabilities. Meanwhile, a series of lightweight specialists are crafted for downstream tasks with low computational cost. In the collaborative inference stage, we introduce two cooperative mechanisms, Mixture-of-Expert Diagnosis and Retrieval-Augmented Diagnosis, to harvest the generalist's in-context learning abilities alongside the specialists' domain expertise. For a comprehensive evaluation, we curate a large-scale benchmark featuring 28 datasets and about 250,000 images. Extensive results demonstrate that MedDr consistently outperforms state-of-the-art GFMs on downstream datasets. Furthermore, GSCo exceeds both GFMs and specialists across all out-of-domain disease diagnosis datasets. These findings indicate a significant paradigm shift in the application of GFMs, transitioning from separate models for specific tasks to a collaborative approach between GFMs and specialists, thereby advancing the frontiers of generalizable AI in medicine.

Motivation & Objective

• To address the scarcity of high-quality, diverse medical image-text datasets for training generalist vision-language models.
• To improve the generalization capability of medical foundation models, especially for rare or unseen diseases.
• To develop a unified generalist model capable of handling multiple medical imaging modalities, including radiology, pathology, dermatology, retinography, and endoscopy.
• To enhance inference reliability through a retrieval-augmented medical diagnosis strategy that leverages external knowledge.
• To demonstrate the effectiveness of Retrieval-Augmented Generation (RAG) in medical vision-language models, a novel application in this domain.

Proposed method

• Proposes a diagnosis-guided bootstrapping method that uses a pre-trained vision-language model to generate detailed medical reports (findings and conclusions) from image-label pairs, ensuring textual consistency with visual content.
• Constructs a large-scale, multi-modal medical vision-language dataset by leveraging label-level annotations from existing medical image datasets, enriching textual descriptions beyond simple captions.
• Trains MedDr, a generalist foundation model on the constructed dataset, enabling zero-shot or few-shot adaptation across diverse medical imaging tasks and modalities.
• Introduces a retrieval-augmented inference strategy that retrieves the top-k most similar images and their annotations during inference to guide and refine model predictions.
• Uses a simple similarity-based retrieval mechanism (e.g., CLIP embeddings) to retrieve relevant examples without fine-tuning, improving robustness and accuracy, especially for rare conditions.
• Employs voting or concatenation of retrieved reports to generate final predictions, enhancing reliability and reducing hallucination in medical report generation and diagnosis.

Experimental results

Research questions

• RQ1 Can diagnosis-guided data generation from image-label pairs produce higher-quality and more consistent vision-language datasets than text-only methods?
• RQ2 Can a generalist foundation model trained on such a dataset generalize across diverse medical imaging modalities, including radiology, pathology, dermatology, retinography, and endoscopy?
• RQ3 Does retrieval-augmented inference significantly improve the accuracy and generalization of a medical vision-language model, especially for rare or unseen diseases?
• RQ4 Is the application of Retrieval-Augmented Generation (RAG) effective and beneficial in the context of medical foundation models, particularly when no fine-tuning is applied?
• RQ5 How does the performance of a generalist model compare to specialist models on rare disease detection, especially when enhanced with retrieval-augmented inference?

Key findings

• MedDr achieves 87.8% accuracy on the BreastMNIST dataset, surpassing the specialist model’s 86.3% accuracy, demonstrating strong generalization even on out-of-distribution data.
• With retrieval-augmented inference, MedDr achieves 95.5% accuracy on the BloodMNIST dataset, which was not in the training set, highlighting its robustness to rare diseases.
• On the MIMIC-CXR benchmark, MedDr generates more comprehensive reports that include both normal and abnormal findings, unlike RadFM, which often defaults to normal status descriptions.
• The retrieval-augmented strategy improves performance across all evaluated tasks—visual question answering, medical report generation, and image diagnosis—outperforming both MedDr without retrieval and Med-Flamingo with retrieval.
• The voting-based retrieval strategy yields higher metrics than baseline methods on multiple datasets, indicating high-quality and relevant retrieved examples.
• Qualitative analysis shows that MedDr generates more accurate and contextually rich reports than specialist models like RadFM, particularly in identifying and describing abnormal findings.