Causal Frameworks and Feature Discrepancy Loss: Addressing Data Scarcity and Enhancing Medical Image Segmentation
Keywords: causal reasoning, bioemdical image segmentation, data dilemma
Abstract: Data scarcity poses a significant challenge for deep learning models in medical imaging, particularly for training and generalization. Previous studies have demonstrated the efficacy of data pooling from various sources, facilitating the analysis of weak but significant correlations between imaging data and disease incidence. This approach is often constrained by strict data-sharing protocols among institutions, resulting in models reliant on external data sources. In this work, we address the issue of data scarcity by leveraging the available data for segmentation tasks across various medical imaging modalities. Based on our observation that samples with minimal foreground-background feature differences often demonstrate inadequate segmentation performance, we propose a causal-inspired foreground-background feature discrepancy penalty function, which improves feature separation and alleviates segmentation difficulties caused by homogeneous pixel distributions. The proposed feature discrepancy loss is mathematically grounded, with a lower bound defined by the negative logarithm of the Dice coefficient, suggesting that increased feature separation correlates with improved Dice scores. To further validate our approach, we introduce a novel ultrasound dataset for triple-negative breast cancer (TNBC), and we evaluate the method across three state-of-the-art segmentation architectures to demonstrate competitive performance. In addition, the results highlight the robustness of our method in mitigating performance decrease due to distribution shifts when new, differently distributed data batches are introduced.
Primary Area: interpretability and explainable AI
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.
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: 11695
Loading