Abstract
Light-sheet microscopy (LSM) can image cleared whole organs and organisms at cellular resolution, but supervised segmentation models often lose accuracy when biological structure, staining protocol, density, or morphology differs from the annotated training set. To this end, the MICCAI 2026 SELMA3D Challenge aims to benchmark selfsupervised learning (SSL) for 3D LSM image segmentation by roviding
a large collection of unannotated 3D LSM human and mice brain image data and relevant annotated cropped patches. Here, we present our twostage teacher-student SSL solution for SELMA3D, based on pretraining a 3D Primus encoder with DINOv2, followed by fine-tuning its exported student encoder with nnU-Net on the provided annotated patches. The SELMA3D setting spans isolated and contiguous structures across sparse and dense regimes, and we compare raw min–max scaling, and intensity thresholding to adapt SSL to 16-bit fluorescence data. Five-fold crossvalidation yielded a i) global Dice Similarity Coefficient of 83.00±2.15, ii) Panoptic Quality of 52.8±3.9 and 57.6±7.2 for dense and sparse isolated structures, respectively, and iii) centerline Dice of 94.8±2.8 and 79.5±3.7 for the dense and sparse contiguous structures, respectively. These results demonstrate promising cross-validated performance of the Primus-DINOv2–nnU-Net pipeline using a computationally constrained subset (6.85%) of the available unannotated data, while preserving the reproducible downstream training framework provided by nnU-Net.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/BrainWorks_031.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=~Spyridon_Bakas3
BibTex
@InProceedings{ThaSid_Selfsupervised_MICCAISAT2026,
author = { Thakur, Siddhesh AND Reddy, Sai Prajwal AND Malec, Sylwia AND Poudel, Chetan AND Garyfallidis, Eleftherios AND Bakas, Spyridon},
title = { { Self-supervised Learning for 3D Light-Sheet Microscopy Image Segmentation based on DINO-v2 and nnU-Net } },
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}
}