Multi-instance Causal Representation Learning-based Network for Glioma Grading

Hanbing Zhang, Guohua Zhao, Yusong Lin
Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, PMLR 278:94-99, 2025.

Abstract

Gliomas are the most common primary intracranial malignant tumors, characterized by high heterogeneity, recurrence, and mortality. Accurate grading is essential for treatment planning and prognosis assessment. MRI, as a non-invasive modality, is widely used, but traditional diagnosis depends on expert experience, leading to subjectivity and inefficiency. AI-based automatic grading has made progress, yet challenges persist due to tumor boundary ambiguity, structural heterogeneity, and the “black box" nature of AI models, limiting robustness, generalization, and interpretability. To address these issues, this study proposes a multi-instance causal representation learning-based network for glioma grading (MCRNet). MCRNet employs multi-instance learning to aggregate MRI slice features, effectively handling tumor heterogeneity. The causal-aware attention mechanism (CAAM) and causal-aware dynamic aggregation mechanism (CDAM) enhance feature selection and aggregation efficiency. Evaluated on BraTS2020 and a private clinical dataset, MCRNet improves robustness, generalization, and interpretability. It minimizes the performance gap between validation and test sets, reducing the AUC difference by up to 3.21% compared to existing methods, demonstrating its potential for reliable clinical application.

Cite this Paper

BibTeX
@InProceedings{pmlr-v278-zhang25b, title = {Multi-instance Causal Representation Learning-based Network for Glioma Grading}, author = {Zhang, Hanbing and Zhao, Guohua and Lin, Yusong}, booktitle = {Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing}, pages = {94--99}, year = {2025}, editor = {Zeng, Nianyin and Pachori, Ram Bilas and Wang, Dongshu}, volume = {278}, series = {Proceedings of Machine Learning Research}, month = {25--27 Apr}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v278/main/assets/zhang25b/zhang25b.pdf}, url = {https://proceedings.mlr.press/v278/zhang25b.html}, abstract = {Gliomas are the most common primary intracranial malignant tumors, characterized by high heterogeneity, recurrence, and mortality. Accurate grading is essential for treatment planning and prognosis assessment. MRI, as a non-invasive modality, is widely used, but traditional diagnosis depends on expert experience, leading to subjectivity and inefficiency. AI-based automatic grading has made progress, yet challenges persist due to tumor boundary ambiguity, structural heterogeneity, and the “black box" nature of AI models, limiting robustness, generalization, and interpretability. To address these issues, this study proposes a multi-instance causal representation learning-based network for glioma grading (MCRNet). MCRNet employs multi-instance learning to aggregate MRI slice features, effectively handling tumor heterogeneity. The causal-aware attention mechanism (CAAM) and causal-aware dynamic aggregation mechanism (CDAM) enhance feature selection and aggregation efficiency. Evaluated on BraTS2020 and a private clinical dataset, MCRNet improves robustness, generalization, and interpretability. It minimizes the performance gap between validation and test sets, reducing the AUC difference by up to 3.21% compared to existing methods, demonstrating its potential for reliable clinical application.} }
Endnote
%0 Conference Paper %T Multi-instance Causal Representation Learning-based Network for Glioma Grading %A Hanbing Zhang %A Guohua Zhao %A Yusong Lin %B Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing %C Proceedings of Machine Learning Research %D 2025 %E Nianyin Zeng %E Ram Bilas Pachori %E Dongshu Wang %F pmlr-v278-zhang25b %I PMLR %P 94--99 %U https://proceedings.mlr.press/v278/zhang25b.html %V 278 %X Gliomas are the most common primary intracranial malignant tumors, characterized by high heterogeneity, recurrence, and mortality. Accurate grading is essential for treatment planning and prognosis assessment. MRI, as a non-invasive modality, is widely used, but traditional diagnosis depends on expert experience, leading to subjectivity and inefficiency. AI-based automatic grading has made progress, yet challenges persist due to tumor boundary ambiguity, structural heterogeneity, and the “black box" nature of AI models, limiting robustness, generalization, and interpretability. To address these issues, this study proposes a multi-instance causal representation learning-based network for glioma grading (MCRNet). MCRNet employs multi-instance learning to aggregate MRI slice features, effectively handling tumor heterogeneity. The causal-aware attention mechanism (CAAM) and causal-aware dynamic aggregation mechanism (CDAM) enhance feature selection and aggregation efficiency. Evaluated on BraTS2020 and a private clinical dataset, MCRNet improves robustness, generalization, and interpretability. It minimizes the performance gap between validation and test sets, reducing the AUC difference by up to 3.21% compared to existing methods, demonstrating its potential for reliable clinical application.
APA
Zhang, H., Zhao, G. & Lin, Y.. (2025). Multi-instance Causal Representation Learning-based Network for Glioma Grading. Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, in Proceedings of Machine Learning Research 278:94-99 Available from https://proceedings.mlr.press/v278/zhang25b.html.

Related Material