Abstract
Medical imaging datasets often contain only few samples due to acquisition costs and privacy constraints while potentially holding clinically valuable information. Standard approaches such as transfer learning and data augmentation also require a minimum amount of data and offer only partial solutions. The question of how to build generalizable models from only few samples remains unsolved. Neuro-symbolic methods that combine data-driven and symbolic reasoning are promising for small-data learning, but remain underinvestigated in the clinical domain. We address this gap by proposing a framework integrating clinical knowledge with three different mechanisms: knowledge-based feature selection, a logic-based loss, and knowledge-guided adversarial training. As a proof-of-concept, we demonstrate our approach in a binary cirrhosis classification task on a multi-parametric MRI dataset of 19 patients, encoding clinical rules via DL2 both as loss and as knowledge-guided adversarial perturbations on a shallow Multi-Layer Perceptron. We find consistent improvement from the knowledge-based feature selection over a radiomics-based approach, and small gains from adding the logic-based loss and the adversarial training, though with high instability across folds. The results suggest that explicit domain knowledge can complement data-centric techniques when sample sizes preclude conventional deep learning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_062.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MLMI_062_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=hfdjKr0Pxy
BibTex
@InProceedings{MerMar_Knowledgeinjection_MICCAISAT2026,
author = { Mertens, Marlene AND Pfister, Felix AND Höhme, Stefan},
title = { { Knowledge-injection for small-cohort medical image learning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}