HiPath: Hierarchical Vision-Language Alignment for Structured Pathology Report Prediction

A research team has introduced HiPath, a lightweight vision-language framework designed to predict structured pathology reports directly from slide images, achieving 68.9% strict accuracy and a 97.3% safety rate in tests across three Chinese hospitals [1]. The framework, detailed in a preprint revised in June 2026, is built on frozen UNI2 and Qwen3 backbones and adds only 15 million trainable parameters through three modules: a Hierarchical Patch Aggregator for multi-image encoding, Hierarchical Contrastive Learning for cross-modal alignment, and Slot-based Masked Diagnosis Prediction for generating structured diagnoses [1][2]. The model was trained on 749,000 real-world pathology cases [1]. Unlike existing pathology vision-language models that typically output flat labels or free-form text, HiPath treats structured report prediction as its central objective [4]. The system predicts typed diagnostic slots by matching against a closed vocabulary in the frozen language model embedding space, requiring only frozen vocabulary embeddings and visual features at inference time [4]. On internal evaluation, HiPath reached 74.7% clinically acceptable accuracy alongside its 68.9% strict accuracy [1]. Vision-only classifiers using the same frozen visual features plateaued at 30–32%, and a non-hierarchical variant with identical text access trailed by 12.9 percentage points, according to the paper [4]. Cross-hospital testing—training on data from two hospitals and evaluating on a third—showed a 3.4 percentage-point drop in strict accuracy while the safety rate remained at 97.1% [1][2]. The authors, led by corresponding author Ruicheng Yuan, have released code and an evaluation protocol on GitHub [4]. The work contributes to a growing body of research on automated histopathology report generation. A separate modular framework proposed in earlier literature paired a frozen UNI foundation model with a lightweight Transformer decoder and BioGPT tokenization, adding a retrieval-based verification step that replaced generated reports with ground-truth references when high-similarity matches were found [7]. That system similarly emphasized computational efficiency without sacrificing diagnostic accuracy [7].