Abstract
Effective representation learning in medical imaging is challenging when both labeled data and computational resources are limited. Feature Learning from Image Markers (FLIM) lets users design efficient convolutional encoders by drawing markers on discriminative regions of a few images; the markers act as weak supervision, and filter coefficients are learned directly from feature patches centered at marker pixels without backpropagation. However, reliance on the user affects reproducibility and scalability. Replacing markers with superpixel centers removes the user from the loop but inflates filter count proportionally to training images and superpixels. We propose Superpixel Filter Learning (SPiFiL), which learns filters from superpixel centers and retains only the most discriminative ones via intra- and inter-class rankings. To showcase robustness, we select three complementary parasitology classification datasets, a critical problem in developing countries that requires an effective solution under low-resource settings. By learning encoder filters from one image per class, SPiFiL can reduce the filter count by up to $25\times$ while achieving competitive performance and improvements of up to $+0.032$ in Cohen’s kappa and $+9.2\%$ in F1-score (using a linear SVM classifier). Compared with strong baselines, SPiFiL achieves competitive performance using only a 0.0525M-parameter encoder against models ranging from 1.2M (SqueezeNet) to 134.3M (VGG16), whether trained from scratch or pretrained on ImageNet, highlighting its potential for scalable, low-resource medical imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_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=aIE0gr3pxm
BibTex
@InProceedings{SalFel_Representation_MICCAISAT2026,
author = { Salvagnini, Felipe C. R. AND Miranda, Maria A. K. AND Soares, Gilson J. AND da silva, Mateus O. AND Santos, Cid AND Neto, Jeova F. S. R. AND Guimarães, Silvio J. F. AND Marques, Oge AND Falcão, Alexandre X.},
title = { { Representation Learning from Superpixels: Lightweight CNN Encoders Under Data Scarcity } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}
