Abstract
Lightweight convolutional neural networks (CNNs) are essential for deploying AI-assisted histopathology-based breast cancer detection (BCD) in computationally constrained healthcare environments. Knowledge distillation (KD) is a natural compression strategy for this setting, but its effectiveness is highly sensitive to the loss-balancing coefficient α and distillation temperature T. In many BCD systems, these hyperparameters are selected manually or via coarse grid search, making optimization difficult to reproduce and limiting practical deployment. We formulate (α, T) selection as a black-box optimization problem and propose a simulated annealing (SA) strategy to automatically tune KD for BCD on the BreakHis dataset. Using validation AUC as the objective, SA searches a continuous (α, T) space under an explicit schedule and stopping criteria, and outperforms fixed, random, and grid search baselines under a matched tuning budget. The SA-tuned student achieves a test AUC of 0.9848, approaching the teacher (AUC 0.9858) while using approximately 116× fewer parameters. These results demonstrate that reproducible KD hyperparameter optimization can facilitate the deployment of accurate lightweight BCD models in resource-constrained healthcare settings. Our code is available at https://github.com/falmatamkn/SA-KD-BreastCancer.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_025.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=afw1uk3s4D
BibTex
@InProceedings{ModFal_ResourceEfficient_MICCAISAT2026,
author = { Modu, Falmata AND Prasad, Rajesh AND Aliyu, Farouq},
title = { { Resource-Efficient Knowledge Distillation via Simulated Annealing for Lightweight Breast Cancer Detection } },
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}
}