Abstract
Geometry-derived biomarkers are downstream of lesion segmentation, yet their reliability is often assumed once a mask is available. We present a lesion-level perturbation audit for multiple-sclerosis lesion geometry that separates three distinct questions: whether a perturbed lesion remains measurable, whether a valid perturbation changes a downstream geometric decision, and whether spectral descriptors themselves remain stable. A paediatric-onset MS reference partition was fitted from 940 PediMS lesions. Its two-cluster resolution is supported by bootstrap stability (mean Jaccard 0.984 for k=2 versus 0.814 for k=3) and leave-one-patient-out reproducibility (ARI 0.997 \pm 0.008). In the primary direct audit of 400 volume-stratified MSLesSeg-2024 lesions, one-voxel erosion, opening, closing and dilation yielded valid-mask rates of 55.3–97.8\% and conditional assignment agreement of 97.4–98.1\%. Treating disappearance itself as a failure gives a more conservative 54.0–95.3\% ``valid-and-unchanged’’ rate. To evaluate the full morphometric-plus-spectral feature space, we also conducted an independent complete-vector audit on 500 real FLAIR lesion components: after 12 binary-mask perturbations, the mesh and full geometry vector, including Shape-DNA, were recomputed for every valid mask. Closing preserved 100% of valid assignments and stochastic boundary dropout/addition preserved 97.8%/96.4%, while dilation was substantially less stable (63.4-80.2%). These results support reliability-aware reporting of lesion geometry while avoiding overinterpretation of a small fragile-lesion subset.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BrainWorks_012.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Joohi_Chauhan1
BibTex
@InProceedings{MitPra_Direct_MICCAISAT2026,
author = { Mittal, Prateek AND Singh, Prabal Pratap AND Chauhan, Joohi},
title = { { Direct Mask-Perturbation Auditing for Geometry-Derived Brain Lesion Biomarkers } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17254},
month = {pending},
page = {pending}
}
