Abstract
We present our submission to the FU_Biometry challenge (MICCAI 2026), a unified multi-task two-dimensional ultrasound landmark-detection problem spanning nine task identifiers across prenatal, intrapartum, and cardiac ultrasound, evaluated by landmark mean radial error (MRE) and derived clinical-parameter mean absolute error (MAE). We construct a leak-free, group-aware cross-validation (CV) harness that provides a reliable offline proxy for the hidden validation score, and use it to systematically evaluate a set of inference-time and training-time interventions applied to the organizers’ DINOv2 multi-task heatmap-regression baseline: sub-pixel heatmap decoding (converting a predicted heatmap into a continuous pixel coordinate rather than the baseline’s quantized argmax), multi-scale test-time augmentation, photometric and geometric training augmentation, a corrected learning-rate schedule, task-specific heatmap resolution, and multi-seed ensembling. Each intervention is retained only if it improves the de-leaked CV task-mean MRE without degrading performance on any individual task. Stacking all retained interventions reduces the CV task-mean MRE from 37.5,px to 25.48,px and the official validation-phase nine-task mean MRE from 38.88,px to 24.28,px, a relative reduction of 37.5%. We additionally report qualitative best- and worst-case predictions on held-out validation folds to characterize where the proposed recipe succeeds and where residual error remains concentrated.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/FoundUS_100.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/forum?id=LKW9RPveuI
BibTex
@InProceedings{AgrHar_MultiTask_MICCAISAT2026,
author = { Agrawal, Harshit},
title = { { Multi-Task Ultrasound Biometry Landmark Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17261},
month = {pending},
page = {pending}
}
