Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Direct Preference Optimization for Suppressing Hallucinated Prior Exams in Radiology Report Generation

Oishi Banerjee, Hong-Yu Zhou, Subathra Adithan, Stephen Kwak, Kay Wu, Pranav Rajpurkar

TL;DR

This work tackles hallucinations in radiology report generation by applying Direct Preference Optimization (DPO) to suppress references to prior exams in chest X-ray reports. It introduces standard and weighted DPO losses, constructs a GPT-4–edited MIMIC-CXR dataset to enable targeted behavior suppression, and demonstrates that DPO can reduce references to prior exams by about 3.2x–4.8x while preserving clinical accuracy. The study provides a first medical VLM application of DPO, highlighting data- and compute-efficient benefits and offering practical guidance for deploying alignment techniques in radiology. The findings suggest that carefully weighted DPO can mitigate harmful model behaviors without sacrificing essential clinical information, potentially reducing radiologist workload and patient risk.

Abstract

Recent advances in generative vision-language models (VLMs) have exciting potential implications for AI in radiology, yet VLMs are also known to produce hallucinations, nonsensical text, and other unwanted behaviors that can waste clinicians' time and cause patient harm. Drawing on recent work on direct preference optimization (DPO), we propose a simple method for modifying the behavior of pretrained VLMs performing radiology report generation by suppressing unwanted types of generations. We apply our method to the prevention of hallucinations of prior exams, addressing a long-established problem behavior in models performing chest X-ray report generation. Across our experiments, we find that DPO fine-tuning achieves a 3.2-4.8x reduction in lines hallucinating prior exams while maintaining model performance on clinical accuracy metrics. Our work is, to the best of our knowledge, the first work to apply DPO to medical VLMs, providing a data- and compute- efficient way to suppress problem behaviors while maintaining overall clinical accuracy.

Direct Preference Optimization for Suppressing Hallucinated Prior Exams in Radiology Report Generation

TL;DR

This work tackles hallucinations in radiology report generation by applying Direct Preference Optimization (DPO) to suppress references to prior exams in chest X-ray reports. It introduces standard and weighted DPO losses, constructs a GPT-4–edited MIMIC-CXR dataset to enable targeted behavior suppression, and demonstrates that DPO can reduce references to prior exams by about 3.2x–4.8x while preserving clinical accuracy. The study provides a first medical VLM application of DPO, highlighting data- and compute-efficient benefits and offering practical guidance for deploying alignment techniques in radiology. The findings suggest that carefully weighted DPO can mitigate harmful model behaviors without sacrificing essential clinical information, potentially reducing radiologist workload and patient risk.

Abstract

Recent advances in generative vision-language models (VLMs) have exciting potential implications for AI in radiology, yet VLMs are also known to produce hallucinations, nonsensical text, and other unwanted behaviors that can waste clinicians' time and cause patient harm. Drawing on recent work on direct preference optimization (DPO), we propose a simple method for modifying the behavior of pretrained VLMs performing radiology report generation by suppressing unwanted types of generations. We apply our method to the prevention of hallucinations of prior exams, addressing a long-established problem behavior in models performing chest X-ray report generation. Across our experiments, we find that DPO fine-tuning achieves a 3.2-4.8x reduction in lines hallucinating prior exams while maintaining model performance on clinical accuracy metrics. Our work is, to the best of our knowledge, the first work to apply DPO to medical VLMs, providing a data- and compute- efficient way to suppress problem behaviors while maintaining overall clinical accuracy.
Paper Structure (20 sections, 3 equations, 3 figures, 6 tables)

This paper contains 20 sections, 3 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Main Insights
  3. Related Work
  4. Methods
  5. Direct Preference Optimization
  6. Pretrained Model
  7. Dataset Creation
  8. Removing References to Prior Exams with GPT-4
  9. Training Dataset
  10. Validation Dataset
  11. Test Datasets
  12. DPO Experiments
  13. Evaluation Approach
  14. Results
  15. GPT-4 Annotation Quality
  16. ...and 5 more sections

Figures (3)

  • Figure 1: GPT-4 labels each line in a report based on how much of it depends on a prior exam ("none", "partial", "all") and rewrites "partial" lines to omit references to prior exams.
  • Figure 2: An overview of our pretraining and fine-tuning methods. Models are rewarded for producing prior-free responses (green) and penalized for generating responses referencing prior exams (red). Depending on the DPO weighting scheme, lines unrelated to prior exams may receive less weight ($\gamma$ = .5) or be entirely ignored ($\gamma$ = 0).
  • Figure 3: DPO fine-tuning dramatically reduces the average number of lines referring to hallucinated prior exams (left), while maintaining clinical accuracy (right). Supervised fine-tuning (SFT) improves clinical accuracy but does not meaningfully prevent the hallucination of prior exams. For both metrics, lower scores indicate higher performance.