Title: MedSynapse-V: Bridging Visual Perception and Clinical Intuition via Latent Memory Evolution

Abstract: Precise medical diagnosis depends not merely on static imaging characteristics, but also on the implicit diagnostic memory that specialists instantly access when interpreting images. We identify a critical cognitive disconnect in medical Vision-Language Models (VLMs) stemming from discrete tokenization, which results in quantization errors, the dissipation of long-range information, and a lack of case-adaptive expertise. To address this discrepancy, we introduce MedSynapse-V, a framework designed for the evolution of latent diagnostic memory. This approach mimics the experiential recall of clinicians by dynamically constructing implicit diagnostic memories within the model’s hidden streams.

The process initiates with a Meta Query for Prior Memorization mechanism. Here, learnable probes extract structured priors from an anatomical prior encoder to produce condensed implicit memories. To maintain clinical accuracy, we incorporate Causal Counterfactual Refinement (CCR). This component utilizes reinforcement learning and counterfactual rewards—generated through region-level feature masking—to measure the causal impact of each memory. Consequently, it eliminates redundancies and aligns latent representations with diagnostic reasoning.

This evolutionary cycle concludes with Intrinsic Memory Transition (IMT), a privileged-autonomous dual-branch paradigm. IMT internalizes diagnostic patterns from a teacher branch into a student branch through full-vocabulary divergence alignment. Extensive empirical assessments across various datasets reveal that MedSynapse-V significantly surpasses current state-of-the-art techniques, especially chain-of-thought approaches, in diagnostic accuracy by transferring external expertise into endogenous parameters. The code is available at https://github.com/zhcz328/MedSynapse-V.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC