Abstract
Clinical adoption of medical AI is constrained by annotation cost and computational resources, not only by accuracy. We study benign–malignant lung nodule classification with a lightweight model that needs no pretraining and stays accurate when labeled data are scarce – the setting in which large pretrained 3D networks tend to overfit and are expensive to deploy. The model builds on Massive-Training Artificial Neural Networks (MTANN), patch-wise regressors with only K parameters, and incorporeates supervised contrastive learning through Multi-Level Semantic Functional Pairing (MLSFP). MLSFP forms positive pairs that share a class label, lie close together anatomically, and produce similar teaching-signal responses. On LIDC-IDRI, MLSFP-MTANN outperformed strong convolutional, transformer, and state-space classifiers in both regimes. At full data, it reached AUC, against at most for the deep baselines; with only 5 to 15 labeled images per class, it held about AUC, while those models remained near chance performance () despite having - times more parameters. The way the pairs are built also matters, since naive same-class pairing is unstable and can lower accuracy, whereas MLSFP improved over the no-contrastive MTANN under scarcity ( and AUC at 5 and 10 images per class, paired ) and was insensitive to its pairing thresholds (AUC range over a grid). A small, efficient model with structured supervision can therefore beat much larger pretrained networks when labels are limited.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/EMA_019.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{BilMus_MLSFPCL_MICCAISAT2026,
author = { Billah, Mustain AND Rahmaniar, Wahyu AND Jin, Ze AND Suzuki, Kenji},
title = { { MLSFP-CL: Label- and Compute-Efficient Contrastive Learning for Lung Nodule Classification in Small-Data Settings } },
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}
}