Abstract
Implicit Neural Representations (INRs) encode images as continuous coordinate-based functions and provide an alternative to fixed pixel-grid representations. We study biomedical image classification in INR weight space, where each image is represented by a frozen SIREN and classified through learned probes. Standard probing treats the INR as a black box and only uses its outputs. We extend this idea with lightweight Transformer based equivariant adapters inserted into hidden layers, allowing probes to modulate internal representations while preserving permutation equivariance. On three biomedical datasets, covering chest X-ray, retinal fundus and breast ultrasound images, the proposed adapter-probing approach consistently improves over weight statistics, graph based weight space learning and standard probing. Qualitative analyses further show that learned probes often move outside the original image coordinate domain and that adapters actively reshape INR representations toward task discriminative structures.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/Off_Grid_038.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=1cdLJvH91R
BibTex
@InProceedings{ByrMic_Probing_MICCAISAT2026,
author = { Byra, Michal AND Pregowska, Agnieszka AND Szczepanski, Janusz},
title = { { Probing implicit neural representations for biomedical image classification in weight space } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17277},
month = {pending},
page = {pending}
}
