Abstract
Lesions in moderate-to-severe traumatic brain injury (msTBI) vary enormously in size, number and tissue involvement, which breaks standard neuroimaging pipelines. It also creates a class-imbalance problem that most challenge reports treat as background detail: in the AIMS-TBI 2026 training set, 41.7% of subjects have no lesion at all. We examine how this imbalance shapes a self-configuring nnU-Net trained on a single 16 GB GPU, and serve both challenge tasks from it. The surface-distance metric separating mid-field entries from leading ones tracks a recall tail rather than boundary quality: roughly 7% of scored cases, the outright misses and gross mislocalisations, carry more than half the mean average symmetric surface distance (ASSD), and removing them drops the cross-validation mean from 8.55 mm to 6.19 mm while the median holds near 2.4 mm. The tail is almost entirely a small-lesion effect: the failure rate falls monotonically with lesion volume and reaches zero above the cohort median, confining the deficit to a characterisable subgroup. Three cheap interventions fail to touch it. Lowering the softmax threshold and batch-pooling the Dice loss are reported as negative results with a mechanism; test-time mirroring, which we scored against the organisers’ hidden labels, improves Dice and HD95 slightly while ASSD moves the wrong way. Because surviving lesion volume separates the classes cleanly, a subject-level volume rule reaches a balanced accuracy of 0.846 on the official test set and 0.924 held-out; segmentation reaches a lesion-containing Dice of 0.505. On imbalanced benchmarks, the shape of the failure distribution deserves reporting alongside the aggregate.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AIMS_TBI_002.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=~Geofray_Paul_J1
BibTex
@InProceedings{JGeo_Diagnosing_MICCAISAT2026,
author = { J, Geofray Paul},
title = { { Diagnosing the Recall Tail: A Class-Imbalance Study of nnU-Net for Detection and Segmentation of Heterogeneous Moderate–Severe TBI Lesions } },
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}
}