Abstract
Automated chest X-ray analysis holds significant promise for improving diagnostic access in low- and middle-income countries (LMICs), where radiologist shortages and limited computational infrastructure remain critical barriers to timely care. State-of-the-art deep learning models achieve strong diagnostic accuracy but demand hardware resources incompatible with constrained clinical deployments. We propose an end-to-end pipeline combining knowledge distillation, post-training 8-bit integer (INT8) quantization, and Grad-CAM explainability to produce a lightweight, interpretable model suitable for resource limited settings. A DenseNet-121 teacher is trained on the NIH ChestXray14 dataset (112,120 images, 14 disease labels) and used to distill knowledge into a MobileNetV3-Small student via a combined hard-label and soft-target loss. The student achieves a mean AUC of 0.8059 across 14 thoracic diseases, retaining 99.7% of teacher performance (AUC 0.8081), while reducing model size by 77.9% (28.5 MB to 6.3 MB) and CPU inference latency by 3.3 times (6.0 ms to 1.8 ms). Post-training INT8 quantization further reduces the footprint to 2.0 MB, a 93% reduction from the teacher. The student outperforms the teacher on seven of fourteen disease classes. Grad-CAM analysis demonstrates that the student preserves clinically meaningful saliency patterns relative to the teacher. To our knowledge, this is the first work to jointly evaluate knowledge distillation, quantization, and explainability on ChestX-ray14 with an explicit resource deployment audit framed for LMIC clinical settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_014.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=dw8yQyAicg
BibTex
@InProceedings{IorAon_Efficient_MICCAISAT2026,
author = { Iorumbur, Aondona Moses AND Sanni, Henry Ananyi AND Raymond, Confidence},
title = { { Efficient Explainable AI for Multi-Label Chest Disease Detection in Resource-Constrained Settings via Knowledge Distillation and Quantization } },
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}
}