Hierarchical Corrector of Medical Imaging Descriptions: A Hybrid Approach Combining Domain Lexicons and GraphRAG-Augmented Knowledge Reasoning

This paper introduces MedImgGraphRag-Corrector, a hierarchical hybrid framework that combines a specialized medical imaging dictionary with GraphRAG-enhanced knowledge inference to correct errors in medical imaging reports. While GraphRAG and its medical extension, MedGraphRAG, perform well in long text comprehension and information integration, they have notable shortcomings in word-level error correction—especially in generating replacement word candidates and handling similar diagnostic expressions. Furthermore, the existing Chinese text correction project "pycorrector" is effective for everyday language but falls short in the medical imaging domain.To address these gaps, this paper builds a binary tree linkage structure based on a knowledge graph, linking correction results to reliable imaging descriptions, diagnostic examples, and explanations of relevant vocabulary. A hierarchical correction mechanism is introduced, employing a multi-level screening strategy for error detection, candidate generation, and context consistency checks. For both accuracy and speed, various large language models were evaluated, and deepseek-v3 was ultimately selected. By generating evidence-based corrections, deepseek-v3 effectively minimizes hallucination and boosts overall correction accuracy.