Abstract
Expert histopathology annotation is costly and institution-specific. We study task-training-free colorectal tissue classification using frozen natural-image DINOv2 features and FAISS nearest-neighbour retrieval. The query set requires no labels at inference time; class labels are available only for the retrieval gallery and are used for neighbour voting. We use “task-training-free” to mean no task-specific classifier fitting or backbone updates. A colour-bias audit establishes an 83.33% histogram baseline (red-channel ANOVA: F = 865, p < 0.001). Across five random gallery-query splits, DINOv2 ViT-S/14 achieves 93.96% mean accuracy (SD 1.63%; 95% t-interval ±2.02%), 10.62 pp above the colour baseline. On the primary split (seed 42), accuracy reaches 95.56% (macro-F1 0.9552). Confidence-filtered retrieval (CF-Retrieval, k_CF = 10) reaches 99.24% accuracy (macro-F1 0.9926) at 88.0% coverage using a held-out calibrated threshold (τ = 0.71). On the external-cohort CRC-VAL-HE-7K set, the frozen gallery reaches 86.48% accuracy (macro-F1 0.8373), 16.78 pp above its colour baseline; the same selective threshold reaches 96.70% accuracy at 70.43% coverage. A perceptual-hash audit finds minimal duplication (minimum Hamming distance: 12/64 bits).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DEMI_020.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=cxQ7Er5Guc
BibTex
@InProceedings{JhaPra_TrainingFree_MICCAISAT2026,
author = { Jha, Prabhash Kumar AND Kafley, Saugat AND Jha, Debesh},
title = { { Training-Free Histopathology Classification via Confidence-Filtered DINOv2 Retrieval } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}