Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MedAD-R1: Eliciting Consistent Reasoning in Interpretible Medical Anomaly Detection via Consistency-Reinforced Policy Optimization

Haitao Zhang, Yingying Wang, Jiaxiang Wang, Haote Xu, Hongyang Zhang, Yirong Chen, Yue Huang, Xinghao Ding

TL;DR

MedAD tackles data fragmentation and reasoning deficiencies in medical anomaly detection with large multimodal models. It introduces the MedAD-38K benchmark, featuring multi-center, multi-modal data with structured VQA and explicit diagnostic CoT annotations, and a two-stage training pipeline that includes Cognitive Injection via SFT and Consistency GRPO (Con-GRPO) to enforce coherent reasoning. Con-GRPO employs a composite reward over formatting, accuracy, and reasoning-consistency, leveraging a memory-efficient group-relative policy optimization without a value network. On MedAD-38K, MedAD-R1, a lightweight 3B model, achieves state-of-the-art accuracy (e.g., 85.15% overall) and delivers transparent, verifiable diagnostic reasoning, significantly outperforming both domain-specific and general-purpose baselines. This work advances clinically trustworthy AI by coupling high diagnostic performance with interpretable, logically grounded reasoning in MedAD.

Abstract

Medical Anomaly Detection (MedAD) presents a significant opportunity to enhance diagnostic accuracy using Large Multimodal Models (LMMs) to interpret and answer questions based on medical images. However, the reliance on Supervised Fine-Tuning (SFT) on simplistic and fragmented datasets has hindered the development of models capable of plausible reasoning and robust multimodal generalization. To overcome this, we introduce MedAD-38K, the first large-scale, multi-modal, and multi-center benchmark for MedAD featuring diagnostic Chain-of-Thought (CoT) annotations alongside structured Visual Question-Answering (VQA) pairs. On this foundation, we propose a two-stage training framework. The first stage, Cognitive Injection, uses SFT to instill foundational medical knowledge and align the model with a structured think-then-answer paradigm. Given that standard policy optimization can produce reasoning that is disconnected from the final answer, the second stage incorporates Consistency Group Relative Policy Optimization (Con-GRPO). This novel algorithm incorporates a crucial consistency reward to ensure the generated reasoning process is relevant and logically coherent with the final diagnosis. Our proposed model, MedAD-R1, achieves state-of-the-art (SOTA) performance on the MedAD-38K benchmark, outperforming strong baselines by more than 10\%. This superior performance stems from its ability to generate transparent and logically consistent reasoning pathways, offering a promising approach to enhancing the trustworthiness and interpretability of AI for clinical decision support.

MedAD-R1: Eliciting Consistent Reasoning in Interpretible Medical Anomaly Detection via Consistency-Reinforced Policy Optimization

TL;DR

MedAD tackles data fragmentation and reasoning deficiencies in medical anomaly detection with large multimodal models. It introduces the MedAD-38K benchmark, featuring multi-center, multi-modal data with structured VQA and explicit diagnostic CoT annotations, and a two-stage training pipeline that includes Cognitive Injection via SFT and Consistency GRPO (Con-GRPO) to enforce coherent reasoning. Con-GRPO employs a composite reward over formatting, accuracy, and reasoning-consistency, leveraging a memory-efficient group-relative policy optimization without a value network. On MedAD-38K, MedAD-R1, a lightweight 3B model, achieves state-of-the-art accuracy (e.g., 85.15% overall) and delivers transparent, verifiable diagnostic reasoning, significantly outperforming both domain-specific and general-purpose baselines. This work advances clinically trustworthy AI by coupling high diagnostic performance with interpretable, logically grounded reasoning in MedAD.

Abstract

Medical Anomaly Detection (MedAD) presents a significant opportunity to enhance diagnostic accuracy using Large Multimodal Models (LMMs) to interpret and answer questions based on medical images. However, the reliance on Supervised Fine-Tuning (SFT) on simplistic and fragmented datasets has hindered the development of models capable of plausible reasoning and robust multimodal generalization. To overcome this, we introduce MedAD-38K, the first large-scale, multi-modal, and multi-center benchmark for MedAD featuring diagnostic Chain-of-Thought (CoT) annotations alongside structured Visual Question-Answering (VQA) pairs. On this foundation, we propose a two-stage training framework. The first stage, Cognitive Injection, uses SFT to instill foundational medical knowledge and align the model with a structured think-then-answer paradigm. Given that standard policy optimization can produce reasoning that is disconnected from the final answer, the second stage incorporates Consistency Group Relative Policy Optimization (Con-GRPO). This novel algorithm incorporates a crucial consistency reward to ensure the generated reasoning process is relevant and logically coherent with the final diagnosis. Our proposed model, MedAD-R1, achieves state-of-the-art (SOTA) performance on the MedAD-38K benchmark, outperforming strong baselines by more than 10\%. This superior performance stems from its ability to generate transparent and logically consistent reasoning pathways, offering a promising approach to enhancing the trustworthiness and interpretability of AI for clinical decision support.
Paper Structure (20 sections, 8 equations, 4 figures, 3 tables)

This paper contains 20 sections, 8 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Medical Anomaly Detection
  4. Large Multimodal Models in Medicine
  5. Reasoning via CoT and RL
  6. Methodology
  7. The MedAD-38K Benchmark
  8. The MedAD-R1 Framework
  9. Cognitive Injection via SFT
  10. Reasoning Reinforcement via Con-GRPO
  11. Reward Function Design
  12. Policy Optimization with Con-GRPO
  13. Experiments
  14. Experimental Setup
  15. Dataset and Evaluation Metrics.
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Comparison of MedAD-R1 against traditional methods and general-purpose LMMs. Traditional methods yield uninterpretable anomaly scores, while general models can fail or produce diagnoses inconsistent with their reasoning. MedAD-R1 provides a transparent and logically coherent diagnostic process, enhancing both interpretability and clinical reliability.
  • Figure 2: Illustrative examples of the five core VQA tasks within the MedAD-38K benchmark. Each task is structured in a multiple-choice format to facilitate standardized evaluation.
  • Figure 3: An overview of our entire pipeline. Top: The construction process for the MedAD-38K benchmark, detailing the automated CoT generation using large models followed by a rigorous manual verification phase. Bottom: The two-stage training architecture of MedAD-R1, which starts with Supervised Fine-tuning to align the model, followed by our Con-GRPO algorithm to reinforce logical reasoning.
  • Figure 4: The five-region spatial grid used for lesion localization in the MedAD-38K benchmark.