A Multitask Framework for Label Refinement and Lesion Segmentation in Clinical Brain Imaging

A Multitask Framework for Label Refinement and Lesion Segmentation in Clinical Brain Imaging

Medical Image Learning with Limited and Noisy Data

10.1007/978-3-031-44917-8_6

http://dx.doi.org/10.1007/978-3-031-44917-8_6

Yang Yu, Jiahao Wang, Ashish Jith Sreejith Kumar, Bryan Tan, Navya Vanjavaka, Nurul Hafidzah Rahim, Alistair Koh, Shaheen Low, Yih Yian Sitoh, Hanry Yu, Pavitra Krishnaswamy, Ivan Ho Mien

07 October 2023

Yu, Y., Wang, J., Sreejith Kumar, A. J., Tan, B., Vanjavaka, N., Rahim, N. H., Koh, A., Low, S., Sitoh, Y. Y., Yu, H., Krishnaswamy, P., & Ho Mien, I. (2023). A Multitask Framework for Label Refinement and Lesion Segmentation in Clinical Brain Imaging. Lecture Notes in Computer Science, 60–70. https://doi.org/10.1007/978-3-031-44917-8_6

Abstract. Supervised deep learning methods offer the potential for automating lesion segmentation in routine clinical brain imaging, but performance is dependent on label quality. In practice, obtaining high quality labels from experienced annotators for large-scale datasets is not always feasible, while noisy labels from less experienced annotators are often available. Prior studies focus on either label refinement methods or on learning to segment with noisy labels, but there has been little work on integrating these approaches within a unified framework. To address this gap, we propose a novel multitask framework for end-to-end noisy-label refinement and lesion segmentation. Our approach minimizes the discrepancy between the refined label and the predicted segmentation mask and is highly customizable for scenarios with multiple sets of noisy labels, incomplete ground truth coverage and/or 2D/3D scans. In extensive experiments on both proprietary and public clinical brain imaging datasets, we demonstrate that our end-to-end framework offers strong performance improvements over prevailing baselines on both label refinement and lesion segmentation. Our proposed framework maintains performance gains over baselines even when ground truth labels are available for only 25–50% of the dataset. Our approach has implications for effective medical image segmentation in settings that are replete with noisy labels but sparse on ground truth annotation.

Publisher Copyright

This research is supported by core funding from: I2R, SERC
Grant Reference no. : NA

This version of the article has been accepted for publication, after peer review and is subject to Springer Nature’s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: http://dx.doi.org/10.1007/978-3-031-44917-8_6

https://oar.a-star.edu.sg/communities-collections/articles/19347

9783031449178

9783031471964

Institute for Infocomm Research