UMamba-ProSSL: Self-Supervised Large-Scale Pretraining with Multi-Task UMamba Advances Prostate Cancer Detection in Biparametric MRI

Syed Farhan Abbas, Michael S. Larsen, Arild Strømsvåg, Tone F. Bathen, Frank Lindseth, Gabriel Kiss, Mattijs Elschot
Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, PMLR 315:4555-4578, 2026.

Abstract

Accurate prostate cancer (PCa) diagnosis is crucial, as it remains one of the leading cause of mortality among men. Although prostate magnetic resonance imaging (MRI) has improved the diagnostic workflow, radiologists still face challenges due to inter-observer variability and limited specificity, leading to both over- and under-diagnosis. Deep learning methods have the potential to support radiologists, but their performance typically depends on large, high-quality labeled datasets that are often scarce and expensive to curate. In contrast, large volumes of unlabeled prostate MRI scans are routinely generated in clinical practice, making self-supervised learning (SSL) a compelling approach to exploit this abundant, untapped resource. However, SSL performance depends strongly on backbone architectures and effective pretext tasks. Moreover, the lack of large-scale standardized benchmarking further limits progress. In this study, we employ a state-of-the-art UMamba for prostate cancer detection and investigate several SSL strategies using a large in-house unlabeled prostate MRI dataset (N=2,431). Among the different pretraining methods, UMamba pretrained with masked autoencoders (MAE) achieved the best downstream performance, with an aggregated mean score of 0.780 (AUROC: 0.905, AP: 0.655) on the large-scale PI-CAI hidden testing set (N=1,000). This performance ranked first on the PI-CAI benchmark leaderboard at the time of evaluation. To further evaluate generalizability, we conducted an evaluation on the out-of-distribution Prostate158 (N=158) dataset, where MAE-pretrained UMamba achieved the best generalization performance, indicating robustness across different clinical centers and imaging protocols. These findings highlighting the strong potential of SSL, particularly MAE combined with UMamba for improving PCa detection accuracy and potentially reducing unnecessary biopsies.

Cite this Paper

BibTeX
@InProceedings{pmlr-v315-abbas26a, title = {UMamba-ProSSL: Self-Supervised Large-Scale Pretraining with Multi-Task UMamba Advances Prostate Cancer Detection in Biparametric MRI}, author = {Abbas, Syed Farhan and Larsen, Michael S. and Str{\o}msv{\aa}g, Arild and Bathen, Tone F. and Lindseth, Frank and Kiss, Gabriel and Elschot, Mattijs}, booktitle = {Proceedings of The 9th International Conference on Medical Imaging with Deep Learning}, pages = {4555--4578}, year = {2026}, editor = {Huo, Yuankai and Gao, Mingchen and Kuo, Chang-Fu and Jin, Yueming and Deng, Ruining}, volume = {315}, series = {Proceedings of Machine Learning Research}, month = {08--10 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v315/main/assets/abbas26a/abbas26a.pdf}, url = {https://proceedings.mlr.press/v315/abbas26a.html}, abstract = {Accurate prostate cancer (PCa) diagnosis is crucial, as it remains one of the leading cause of mortality among men. Although prostate magnetic resonance imaging (MRI) has improved the diagnostic workflow, radiologists still face challenges due to inter-observer variability and limited specificity, leading to both over- and under-diagnosis. Deep learning methods have the potential to support radiologists, but their performance typically depends on large, high-quality labeled datasets that are often scarce and expensive to curate. In contrast, large volumes of unlabeled prostate MRI scans are routinely generated in clinical practice, making self-supervised learning (SSL) a compelling approach to exploit this abundant, untapped resource. However, SSL performance depends strongly on backbone architectures and effective pretext tasks. Moreover, the lack of large-scale standardized benchmarking further limits progress. In this study, we employ a state-of-the-art UMamba for prostate cancer detection and investigate several SSL strategies using a large in-house unlabeled prostate MRI dataset (N=2,431). Among the different pretraining methods, UMamba pretrained with masked autoencoders (MAE) achieved the best downstream performance, with an aggregated mean score of 0.780 (AUROC: 0.905, AP: 0.655) on the large-scale PI-CAI hidden testing set (N=1,000). This performance ranked first on the PI-CAI benchmark leaderboard at the time of evaluation. To further evaluate generalizability, we conducted an evaluation on the out-of-distribution Prostate158 (N=158) dataset, where MAE-pretrained UMamba achieved the best generalization performance, indicating robustness across different clinical centers and imaging protocols. These findings highlighting the strong potential of SSL, particularly MAE combined with UMamba for improving PCa detection accuracy and potentially reducing unnecessary biopsies.} }
Endnote
%0 Conference Paper %T UMamba-ProSSL: Self-Supervised Large-Scale Pretraining with Multi-Task UMamba Advances Prostate Cancer Detection in Biparametric MRI %A Syed Farhan Abbas %A Michael S. Larsen %A Arild Strømsvåg %A Tone F. Bathen %A Frank Lindseth %A Gabriel Kiss %A Mattijs Elschot %B Proceedings of The 9th International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2026 %E Yuankai Huo %E Mingchen Gao %E Chang-Fu Kuo %E Yueming Jin %E Ruining Deng %F pmlr-v315-abbas26a %I PMLR %P 4555--4578 %U https://proceedings.mlr.press/v315/abbas26a.html %V 315 %X Accurate prostate cancer (PCa) diagnosis is crucial, as it remains one of the leading cause of mortality among men. Although prostate magnetic resonance imaging (MRI) has improved the diagnostic workflow, radiologists still face challenges due to inter-observer variability and limited specificity, leading to both over- and under-diagnosis. Deep learning methods have the potential to support radiologists, but their performance typically depends on large, high-quality labeled datasets that are often scarce and expensive to curate. In contrast, large volumes of unlabeled prostate MRI scans are routinely generated in clinical practice, making self-supervised learning (SSL) a compelling approach to exploit this abundant, untapped resource. However, SSL performance depends strongly on backbone architectures and effective pretext tasks. Moreover, the lack of large-scale standardized benchmarking further limits progress. In this study, we employ a state-of-the-art UMamba for prostate cancer detection and investigate several SSL strategies using a large in-house unlabeled prostate MRI dataset (N=2,431). Among the different pretraining methods, UMamba pretrained with masked autoencoders (MAE) achieved the best downstream performance, with an aggregated mean score of 0.780 (AUROC: 0.905, AP: 0.655) on the large-scale PI-CAI hidden testing set (N=1,000). This performance ranked first on the PI-CAI benchmark leaderboard at the time of evaluation. To further evaluate generalizability, we conducted an evaluation on the out-of-distribution Prostate158 (N=158) dataset, where MAE-pretrained UMamba achieved the best generalization performance, indicating robustness across different clinical centers and imaging protocols. These findings highlighting the strong potential of SSL, particularly MAE combined with UMamba for improving PCa detection accuracy and potentially reducing unnecessary biopsies.
APA
Abbas, S.F., Larsen, M.S., Strømsvåg, A., Bathen, T.F., Lindseth, F., Kiss, G. & Elschot, M.. (2026). UMamba-ProSSL: Self-Supervised Large-Scale Pretraining with Multi-Task UMamba Advances Prostate Cancer Detection in Biparametric MRI. Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 315:4555-4578 Available from https://proceedings.mlr.press/v315/abbas26a.html.

Related Material