Hallucination Mitigating for Medical Report Generation

TL;DR

This work refines the input to the LVLM by first utilizing MedCLIP for knowledge retrieval, incorporating relevant lesion fact sentences from a curated knowledge corpus, and introduces a novel purification module to ensure the retrieved knowledge is contextually relevant to the patient's clinical context.

Abstract

In the realm of medical report generation (MRG), the integration of natural language processing has emerged as a vital tool to alleviate the workload of radiologists. Despite the impressive capabilities demonstrated by large vision language models (LVLMs) in understanding natural language, their susceptibility to generating plausible yet inaccurate claims, known as ``hallucinations'', raises concerns-especially in the nuanced and critical field of medical. In this work, we introduce a framework, \textbf{K}nowledge-\textbf{E}nhanced with Fine-Grained \textbf{R}einforced Rewards \textbf{M}edical Report Generation (KERM), to tackle the issue. Our approach refines the input to the LVLM by first utilizing MedCLIP for knowledge retrieval, incorporating relevant lesion fact sentences from a curated knowledge corpus. We then introduce a novel purification module to ensure the retrieved knowledge is contextually relevant to the patient's clinical context. Subsequently, we employ fine-grained rewards to guide these models in generating highly supportive and clinically relevant descriptions, ensuring the alignment of model's outputs with desired behaviors. Experimental results on IU-Xray and MIMIC-CXR datasets validate the effectiveness of our approach in mitigating hallucinations and enhancing report quality.

Figures (6)

• Figure 1: An example of the report generated by the LVLM, where the terms marked in red are hallucinations.
• Figure 2: Overview of KERM. We first retrieve the knowledge from our constructed Knowledge Corpus to enhance the image representation as additional input. During the training period, we employ CheXpert to obtain disease labels, applying penalties to hallucinatory content at both the disease and sentence levels. This reward is then feedback to the LVLM, thereby guiding the model's performance.
• Figure 3: The prompt for generating sentence-level score that scored by GPT-3.5.
• Figure 4: Analysis of the hyperparameter $\mathbf{\alpha}$ with respect to F1 and BLEU-4 on MIMIC-CXR dataset.
• Figure 5: Illustrations of reports from ground truth, ours and Base. For better visualization, different colors highlight different medical terms. The terms marked in red are hallucinations, the terms marked in blue means descriptions included in Ground-Truth but not mentioned in the base model.