Abstract
Foundation Models (FMs) based on the self-supervised DINOv2 architecture have become the standard approach for feature extraction in computational pathology. In these models, embedding size is typically inherited from the FM and treated as a fixed design choice rather than a hyperparameter. However, there is growing evidence that embeddings produced by this class of FMs exist in a space with an intrinsic dimensionality substantially lower than the number of dimensions in the embedding. This suggests that embeddings can be projected into more compact representations, saving substantial computational resources for downstream tasks. This work evaluates three projection techniques (principal component analysis (PCA), uniform manifold approximation and projection (UMAP), and Gaussian random projection (GRP)) via downstream classification and utilises five datasets covering four tissue types (colorectal, breast, skin, thorax), and across four FMs: three histopathology FMs (UNI, UNI2-h, and Hibou-L) and DINOv2 as a general-purpose baseline. Across the 20 dataset–encoder pairs, PCA recovers 99% of full-dimensional classification ROC-AUC at a median target dimension of 9, reducing dense float32 feature storage by 113.8–170.7 relative to 1024-D and 1536-D embeddings. Furthermore, compressing to the estimated intrinsic dimensionality preserves near-parity or surpasses full-dimensional performance for the histopathology FMs, whereas DINOv2 shows greater degradation at lower dimensions.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMA_012.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: Not Available
BibTex
@InProceedings{HosMir_Nine_MICCAISAT2026,
author = { Hossain, Mirza Nasir AND Morrison, David AND Dreiling, Julia AND Harris-Birtill, David},
title = { { Nine Dimensions Are Enough for Patch-Level Classification with Histopathology Foundation Model Embeddings } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}