Abstract
Modern deep learning methods for medical image segmentation require large training datasets and substantial computational power for both training and deployment. This could prevent the use of these methods in domains such as parasitology, where only a few annotated datasets are available, and in regions with limited access to computational resources. Previous work introduces Feature Learning from Image Markers (FLIM) as a methodology for training CNN encoders using only patches extracted from marker annotations of a few representative images. An advantage of this method is that it requires less data and can achieve close to state-of-the-art results with small networks. Despite these advances, prior approaches still rely on users for image selection, marker annotation, and architecture definition, constraining scalability. In this work, we overcome these limitations with a methodology for jointly optimizing the architecture, markers, and hyperparameters, which we call AutoFLIM. The method uses a Genetic Algorithm with a multi-objective formulation that balances segmentation (Dice coefficient) and computational efficiency (FLOPs). Because FLIM training requires no backpropagation, the search explores a far larger space than a grid search or manual design. We evaluate our approach on two parasitology microscopy datasets, demonstrating that competitive segmentation can be achieved with only a few training images, approaching the performance of more computationally intensive approaches in the low-data regime, while using $322-558\times$ fewer parameters and $14-42\times$ fewer FLOPs. The final encoder is built from only 4–5 marked images, showing that few images suffice to train a competitive model; however, the greedy selection step still scores candidates against the annotated pool, so the current annotation budget remains that of the full labeled set. Removing this dependence is the natural next step. The code is available at https://github.com/LIDS-UNICAMP/AutoFLIM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_107.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=34QUVUFTxe
BibTex
@InProceedings{SoaGil_AutoFLIM_MICCAISAT2026,
author = { Soares, Gilson J. AND Salvagnini, Felipe C. R. AND Neto, Jeova F. S. R. AND Falcão, Alexandre X.},
title = { { AutoFLIM: Joint Optimization of Architecture, Markers, and Hyperparameters for Efficient Medical Image Segmentation } },
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}
}