Incentivizing Cardiologist-Like Reasoning in MLLMs for Interpretable Echocardiographic Diagnosis

TL;DR

The paper tackles the challenge of interpretable, cardiologist-like reasoning in multimodal LLMs for echocardiographic diagnosis by introducing Cardiac Reasoning Template (CRT) and CardiacMind. CRT provides a library of stepwise, guideline-grounded diagnostic procedures for 15 complex cardiac diseases to anchor reasoning, while CardiacMind employs a two-stage GRPO reinforcement learning framework with three novel rewards (PQtR, PQlR, ESR) and Template-guided Reasoning Rectification to align model reasoning with clinical workflows. On EchoComplex multiview data and public CardiacNet benchmarks, CardiacMind demonstrates substantial accuracy and reasoning quality gains, with a clinician study showing high alignment with cardiologist logic (93.33%). The approach offers scalable, interpretable echocardiographic reasoning and holds promise for improving clinical trust and diagnostic reliability in real-world settings.

Abstract

Echocardiographic diagnosis is vital for cardiac screening yet remains challenging. Existing echocardiography foundation models do not effectively capture the relationships between quantitative measurements and clinical manifestations, whereas medical reasoning multimodal large language models (MLLMs) require costly construction of detailed reasoning paths and remain ineffective at directly incorporating such echocardiographic priors into their reasoning. To address these limitations, we propose a novel approach comprising Cardiac Reasoning Template (CRT) and CardiacMind to enhance MLLM's echocardiographic reasoning by introducing cardiologist-like mindset. Specifically, CRT provides stepwise canonical diagnostic procedures for complex cardiac diseases to streamline reasoning path construction without the need for costly case-by-case verification. To incentivize reasoning MLLM under CRT, we develop CardiacMind, a new reinforcement learning scheme with three novel rewards: Procedural Quantity Reward (PQtR), Procedural Quality Reward (PQlR), and Echocardiographic Semantic Reward (ESR). PQtR promotes detailed reasoning; PQlR promotes integration of evidence across views and modalities, while ESR grounds stepwise descriptions in visual content. Our methods show a 48% improvement in multiview echocardiographic diagnosis for 15 complex cardiac diseases and a 5% improvement on CardiacNet-PAH over prior methods. The user study on our method's reasoning outputs shows 93.33% clinician agreement with cardiologist-like reasoning logic. Our code will be available.

Figures (12)

• Figure 1: Comparison of CardiacMind with prior methods and overview of the echocardiography diagnosis workflow and the Cardiac Reasoning Template. a) PanEcho holste2025complete predictions versus our model's reasoning process. b) Reasoning process comparison between the latest medical reasoning method xu2025lingshu and our model. c) Summary of the dilated cardiomyopathy (DCM) clinical diagnosis workflow bozkurt2016current. d) Overview of the proposed Cardiac Reasoning Template.
• Figure 2: Overview of CardiacMind. a) Training. CardiacMind uses Group Relative Policy Optimization with three novel rewards and two basic rewards. b) Reward design. We introduce three new rewards that encourage extensive reasoning aligned with standard diagnostic procedures. Elements in Green indicate contents from the CRT template. c) Example of the reasoning process. d) Inference. CardiacMind supports scalable inference through template retrieval and Template-guided Reasoning Rectification (TRR). TRR monitors and corrects steps that deviate the prescribed procedure and it refines the final conclusion.
• Figure 3: Average response length during training. "Basic" uses only accuracy and format rewards and yields short reasoning without detail on complex echocardiography inputs. CardiacMind first preserves a stepwise reasoning structure in the first training stage (epoch 0-1). It is then incentivized to produce detailed stepwise diagnostic reasoning in the second training stage (epoch 1-2).
• Figure 4: Cardiologist preference result on reasoning quality between our method and LingShu xu2025lingshu. We assess reasoning quality using three criteria. "Cardiologists Logic": alignment between the reasoning path and cardiologist diagnostic logic. "Clear and Deductive": clarity and deductive structure of the reasoning process. "View & Measurement Involvement": explicit use of views and measurements in the reasoning process. A higher preference number indicates that cardiologists judged the reasoning to be of higher quality.
• Figure 5: The template structure of Cardiac Reasoning Template.