A Weakly Supervised and Globally Explainable Learning Framework for Brain Tumor Segmentation
Keywords: brain tumor segmentation, weakly supervised learning, explainable learning, counterfactual generation, class association embedding, topological data analysis
Abstract: Brain tumors are a prevalent clinical disease that causes significant suffering for patients. Machine-based segmentation of brain tumors can assist doctors in diagnosis and providing better treatment. However, the complex structure of brain tumors presents a challenge for automatic tumor detection. Deep learning techniques have shown great potential in learning feature representations, but they often require a large number of samples with pixel-level annotations for training for implementing objects segmentation. Additionally, the lack of interpretability in deep learning models hinders their application in medical scenarios. In this paper, we propose a counterfactual generation framework that not only achieves exceptional performance in brain tumor segmentation without the need for pixel-level annotations, but also provides explainability. Our framework effectively separate class-related features from class-unrelated features of the samples, and generate new samples that preserve identity features while altering class attributes by embedding different class-related features. We can accurately identify tumor regions through performing comparison between original abnormal images and generated normal samples which preserve original identity features. We employ topological data analysis for projecting extracted class-related features into a globally explainable class-related manifold. Furthermore, by actively manipulating the generation of images with different class attributes with defined paths, we can provide a more comprehensive and robust explanation of the model. We evaluate our proposed method through experiments conducted on two datasets, which demonstrates superior performance of brain segmentation.
Primary Area: representation learning for 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/2024/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: 9332
Loading