Newshot

Component-Adaptive and Lesion-Level Supervision for Improved Small Structure Segmentation in Brain MRI

Summary

arXiv:2604.08015v5 Announce Type: replace Abstract: Small lesions in brain magnetic resonance imaging (MRI) are easily missed: a segmentation model can reach a high Dice similarity coefficient (DSC) while leaving many of them undetected. We ask whether this omission comes from the network or from the training objective. In multiple sclerosis lesion segmentation, nnU-Net trained with its default Dice and cross-entropy loss produces almost no response to most of the small lesions it misses, so no decision threshold recovers them, and four other architectures trained with the same loss detect no more of them. We therefore change only the objective. CATMIL adds two lesion-level terms to the nnU-Net loss: a Component-Adaptive Tversky (CAT) term, which gives each lesion a weight nearly independent of its size, and a Multiple Instance Learning (MIL) term, which treats each lesion as a bag of voxels and penalizes any lesion in which no voxel is detected. On MSLesSeg, CATMIL raises small-lesion…

Original text(via arXiv cs.CV)

Computer Science > Computer Vision and Pattern Recognition

Title:Component-Adaptive and Lesion-Level Supervision for Improved Small Structure Segmentation in Brain MRI

Abstract:Small lesions in brain magnetic resonance imaging (MRI) are easily missed: a segmentation model can reach a high Dice similarity coefficient (DSC) while leaving many of them undetected. We ask whether this omission comes from the network or from the training objective. In multiple sclerosis lesion segmentation, nnU-Net trained with its default Dice and cross-entropy loss produces almost no response to most of the small lesions it misses, so no decision threshold recovers them, and four other architectures trained with the same loss detect no more of them. We therefore change only the objective. CATMIL adds two lesion-level terms to the nnU-Net loss: a Component-Adaptive Tversky (CAT) term, which gives each lesion a weight nearly independent of its size, and a Multiple Instance Learning (MIL) term, which treats each lesion as a bag of voxels and penalizes any lesion in which no voxel is detected. On MSLesSeg, CATMIL raises small-lesion recall from 0.796 to 0.873 (difference 95\% CI +0.030 to +0.157) and misses 48\% fewer lesions, at comparable DSC. The gain holds on the cross-validation predictions of 47 further patients, for lesions of at least 3~mm, and on a second dataset trained with the same loss weights. CATMIL also detects more small lesions than CC-DiceCE, an instance-level loss that balances lesions within an overlap term (0.873 vs.\ 0.831), indicating that lesions must be supervised not only in proportion but also for their detection. The cost is additional small false-positive components; a component-size filter removes most of them, and at the lesion-wise precision of the default loss CATMIL still detects more small lesions. Code: this https URL.
Comments:v05: This version added results for baselines with CC-DiceCE loss
Subjects:Computer Vision and Pattern Recognition (cs.CV) ; Machine Learning (cs.LG)
Cite as:arXiv:2604.08015 [cs.CV]
(or arXiv:2604.08015v5 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2604.08015 arXiv-issued DOI via DataCite

Submission history

Access Paper:

  • HTML (experimental)

Additional Features

  • Audio Summary

Current browse context:

References & Citations

  • Google Scholar
  • Semantic Scholar

BibTeX formatted citation

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by arXiv cs.CV
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗