Debiased Medical Report Generation with High-Frequency Amplification
Keywords: Medical Report Generation, Debiased Generation, Visual Bias, Textual Bias, Frequency Bias, Fourier Transform, High-pass Filtering
Abstract: In recent years, automated medical report generation (MRG) has gained significant research value for its potential to reduce workload and prevent diagnostic errors. However, generating accurate radiology reports remains challenging due to the prevalence of normal regions in X-ray images and normal descriptions in medical reports. Despite various efforts to address these issues, the definitions of visual bias and textual bias remain unclear and there is still a lack of comprehensive analysis of how these biases affect model behavior.
In this work, we rigorously define and conduct an in-depth examination of visual and textual biases inherent in MRG dataset. Our analysis emphasizes that global patterns, such as normal regions and findings, contribute to visual and textual bias. Further, we discuss how these biases make MRG models especially prone to frequency bias, where models tend to prioritize low-frequency signals that capture global patterns, while neglecting high-frequency signals. To debiase the frequency bias, we propose the high-frequency amplification layer (HAL), aimed at enhancing the model's perceptiveness to fine-grained details. Our extensive experiments show that by amplifying high-frequency signals, HAL reduces both visual and textual biases, leading to improved performance in MRG tasks.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13715
Loading