MedAutoCorrect: Image-Conditioned Autocorrection in Medical Reporting

TL;DR

This work tackles the critical problem of inaccuracies in radiology reports by introducing image-conditioned autocorrection. It develops a two-stage DETECT+CORRECT framework: a token-level error detector leveraging a Vision Transformer and a medical-text encoder, followed by a GPT-2-based error-corrector that uses image features to produce clinically accurate replacements. Error data are generated via synthetic injection across six error categories, enabling robust training despite domain shifts from general Internet data. The approach yields improvements in both token-level detection and retrieval-based report generation, demonstrating its potential as a guardrail to enhance the reliability and safety of automated medical reporting. The framework highlights a practical path toward safer AI-assisted radiology, while acknowledging dataset limitations and the need for careful deployment in clinical contexts.

Abstract

In medical reporting, the accuracy of radiological reports, whether generated by humans or machine learning algorithms, is critical. We tackle a new task in this paper: image-conditioned autocorrection of inaccuracies within these reports. Using the MIMIC-CXR dataset, we first intentionally introduce a diverse range of errors into reports. Subsequently, we propose a two-stage framework capable of pinpointing these errors and then making corrections, simulating an \textit{autocorrection} process. This method aims to address the shortcomings of existing automated medical reporting systems, like factual errors and incorrect conclusions, enhancing report reliability in vital healthcare applications. Importantly, our approach could serve as a guardrail, ensuring the accuracy and trustworthiness of automated report generation. Experiments on established datasets and state of the art report generation models validate this method's potential in correcting medical reporting errors.

Figures (10)

• Figure 1: Overview of our DETECT + CORRECT error-correction method
• Figure 2: Overview of the Proposed Framework for Autocorrecting Radiology Reports. The training phase initiates with separate encoding processes for images and text. The encoded representations are then processed by an error identification module, which utilizes three distinct approaches to detect inaccuracies. Subsequently, a language model is fine-tuned on the image-contextualized reports, where injected errors are represented by [ERROR] tokens. During inference, the model applies the error detection mechanism to localize errors that later are replaced with the masked tokens which are corrected by the error correction mechanism.
• Figure 3: Error injection example. In this example, we automatically introduce errors that fall within the categories of incorrect prediction.
• Figure 4: Qualitative example of medical autocorrection: The top section shows the initial report with errors. The bottom section displays the report after processing by our model, with corrections in green and erroneous terms struck through. This exemplifies the model's capability to identify and rectify inaccuracies within clinical text.
• Figure 5: Overview of varied error types in radiological reports, as altered via GPT-4 prompts. Top-left focuses on false predictions, top-right on mislocations, and the bottom on severity misjudgments, illustrating common error types in clinical radiology and their potential impacts on diagnostic accuracy. Some reports in the dataset consist of a mixture of errors.