crossMoDA Challenge: Evolution of Cross-Modality Domain Adaptation Techniques for Vestibular Schwannoma and Cochlea Segmentation from 2021 to 2023

  • 2025-06-13 18:56:39
  • Navodini Wijethilake, Reuben Dorent, Marina Ivory, Aaron Kujawa, Stefan Cornelissen, Patrick Langenhuizen, Mohamed Okasha, Anna Oviedova, Hexin Dong, Bogyeong Kang, Guillaume SallĂ©, Luyi Han, Ziyuan Zhao, Han Liu, Tao Yang, Shahad Hardan, Hussain Alasmawi, Santosh Sanjeev, Yuzhou Zhuang, Satoshi Kondo, Maria Baldeon Calisto, Shaikh Muhammad Uzair Noman, Cancan Chen, Ipek Oguz, Rongguo Zhang, Mina Rezaei, Susana K. Lai-Yuen, Satoshi Kasai, Chih-Cheng Hung, Mohammad Yaqub, Lisheng Wang, Benoit M. Dawant, Cuntai Guan, Ritse Mann, Vincent Jaouen, Ji-Wung Han, Li Zhang, Jonathan Shapey, Tom Vercauteren
  • 0

Abstract

The cross-Modality Domain Adaptation (crossMoDA) challenge series, initiatedin 2021 in conjunction with the International Conference on Medical ImageComputing and Computer Assisted Intervention (MICCAI), focuses on unsupervisedcross-modality segmentation, learning from contrast-enhanced T1 (ceT1) andtransferring to T2 MRI. The task is an extreme example of domain shift chosento serve as a meaningful and illustrative benchmark. From a clinicalapplication perspective, it aims to automate Vestibular Schwannoma (VS) andcochlea segmentation on T2 scans for more cost-effective VS management. Overtime, the challenge objectives have evolved to enhance its clinical relevance.The challenge evolved from using single-institutional data and basicsegmentation in 2021 to incorporating multi-institutional data and Koos gradingin 2022, and by 2023, it included heterogeneous routine data andsub-segmentation of intra- and extra-meatal tumour components. In this work, wereport the findings of the 2022 and 2023 editions and perform a retrospectiveanalysis of the challenge progression over the years. The observations from thesuccessive challenge contributions indicate that the number of outliersdecreases with an expanding dataset. This is notable since the diversity ofscanning protocols of the datasets concurrently increased. The winning approachof the 2023 edition reduced the number of outliers on the 2021 and 2022 testingdata, demonstrating how increased data heterogeneity can enhance segmentationperformance even on homogeneous data. However, the cochlea Dice score declinedin 2023, likely due to the added complexity from tumour sub-annotationsaffecting overall segmentation performance. While progress is still needed forclinically acceptable VS segmentation, the plateauing performance suggests thata more challenging cross-modal task may better serve future benchmarking.

 

Quick Read (beta)

loading the full paper ...