This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
RoentMod: A Synthetic Chest X-Ray Modification Model to Identify and Correct Image Interpretation Model Shortcuts
Created by
Haebom
Author
Lauren H. Cooke, Matthias Jung, Jan M. Brendel, Nora M. Kerkovits, Borek Foldyna, Michael T. Lu, Vineet K. Raghu
Outline
This paper presents RoentMod, a novel framework for solving the problem of shortcut learning in medical imaging AI models. RoentMod combines a conventional medical image generation model (RoentGen) with an image transformation model to generate medically realistic chest X-ray radiographs (CXRs) that synthesize user-selected lesions while preserving other anatomical features of the original images. In an independent study with radiologists, RoentMod-generated images demonstrated high realism and demonstrated a trend toward shortcut learning in state-of-the-art multi-task and baseline models. Augmenting training data using RoentMod improved the model's ability to discriminate across multiple pathologies, suggesting that RoentMod is a useful tool for enhancing the robustness and interpretability of medical imaging AI models.
Takeaways, Limitations
•
Takeaways:
◦
RoentMod presents an effective method to solve the problem of shortened learning of medical imaging AI models.
◦
Contributes to improving the robustness and interpretability of medical imaging AI models.
◦
We provide a general strategy applicable to various medical image analysis models.
◦
Increase model reliability through counterfactual image editing.
◦
We demonstrated performance improvement through AUC enhancement.
•
Limitations:
◦
The performance improvement effect of RoentMod may vary depending on the dataset and model.
◦
External testing did not observe performance improvements for all pathologies.
◦
It depends on the performance of medical image generation models such as RoentGen.
◦
It may be difficult to ensure perfect realism in synthetic images.
