Bridging the Reliability Gap: INT8 Quantization Effects on Discrimination and Calibration in Medical Imaging

Okan Bilge Ozdemir, Ruowang Li
Proceedings of the 7th Conference on Health, Inference, and Learning, PMLR 333:952-965, 2026.

Abstract

Deploying medical imaging classifiers often requires reduced-precision inference for practical latency and memory budgets, yet the impact of quantization on discrimination and calibration varies across tasks and architectures. We evaluate three public medical imaging datasets (BrainMRI, ChestXray, SkinCancer) and eight ImageNet-pretrained backbones under FP32, FP16, INT8 post-training quantization (PTQ), and INT8 quantization-aware training (QAT). We report macro one-vs-rest ROC-AUC and AUPRC, calibration metrics (ECE, Brier score), and efficiency metrics (throughput, p50 and p99 batch latency) measured on GPU and CPU. FP16 closely matches FP32 across datasets, while INT8-PTQ can introduce substantial and architecture-dependent degradation and calibration shifts. INT8-QAT largely recovers floating-point behavior while enabling integer inference. These results motivate evaluating accuracy, calibration, and efficiency together when selecting quantization strategies for clinical deployment.

Cite this Paper

BibTeX
@InProceedings{pmlr-v333-ozdemir26a, title = {Bridging the Reliability Gap: INT8 Quantization Effects on Discrimination and Calibration in Medical Imaging}, author = {Ozdemir, Okan Bilge and Li, Ruowang}, booktitle = {Proceedings of the 7th Conference on Health, Inference, and Learning}, pages = {952--965}, year = {2026}, editor = {Healey, Elizabeth and Fries, Jason and Pollard, Tom and Tang, Shengpu and Zink, Anna and Hartvigsen, Tom and Agrawal, Monica and Finlayson, Sam and Glicksberg, Benjamin and Beaulieu-Jones, Brett and Wang, Kai and Fontalvo, Daseyra and Sarker, Tasmie and Chen, Irene and Alsentzer, Emily}, volume = {333}, series = {Proceedings of Machine Learning Research}, month = {29--30 Jun}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v333/main/assets/ozdemir26a/ozdemir26a.pdf}, url = {https://proceedings.mlr.press/v333/ozdemir26a.html}, abstract = {Deploying medical imaging classifiers often requires reduced-precision inference for practical latency and memory budgets, yet the impact of quantization on discrimination and calibration varies across tasks and architectures. We evaluate three public medical imaging datasets (BrainMRI, ChestXray, SkinCancer) and eight ImageNet-pretrained backbones under FP32, FP16, INT8 post-training quantization (PTQ), and INT8 quantization-aware training (QAT). We report macro one-vs-rest ROC-AUC and AUPRC, calibration metrics (ECE, Brier score), and efficiency metrics (throughput, p50 and p99 batch latency) measured on GPU and CPU. FP16 closely matches FP32 across datasets, while INT8-PTQ can introduce substantial and architecture-dependent degradation and calibration shifts. INT8-QAT largely recovers floating-point behavior while enabling integer inference. These results motivate evaluating accuracy, calibration, and efficiency together when selecting quantization strategies for clinical deployment.} }
Endnote
%0 Conference Paper %T Bridging the Reliability Gap: INT8 Quantization Effects on Discrimination and Calibration in Medical Imaging %A Okan Bilge Ozdemir %A Ruowang Li %B Proceedings of the 7th Conference on Health, Inference, and Learning %C Proceedings of Machine Learning Research %D 2026 %E Elizabeth Healey %E Jason Fries %E Tom Pollard %E Shengpu Tang %E Anna Zink %E Tom Hartvigsen %E Monica Agrawal %E Sam Finlayson %E Benjamin Glicksberg %E Brett Beaulieu-Jones %E Kai Wang %E Daseyra Fontalvo %E Tasmie Sarker %E Irene Chen %E Emily Alsentzer %F pmlr-v333-ozdemir26a %I PMLR %P 952--965 %U https://proceedings.mlr.press/v333/ozdemir26a.html %V 333 %X Deploying medical imaging classifiers often requires reduced-precision inference for practical latency and memory budgets, yet the impact of quantization on discrimination and calibration varies across tasks and architectures. We evaluate three public medical imaging datasets (BrainMRI, ChestXray, SkinCancer) and eight ImageNet-pretrained backbones under FP32, FP16, INT8 post-training quantization (PTQ), and INT8 quantization-aware training (QAT). We report macro one-vs-rest ROC-AUC and AUPRC, calibration metrics (ECE, Brier score), and efficiency metrics (throughput, p50 and p99 batch latency) measured on GPU and CPU. FP16 closely matches FP32 across datasets, while INT8-PTQ can introduce substantial and architecture-dependent degradation and calibration shifts. INT8-QAT largely recovers floating-point behavior while enabling integer inference. These results motivate evaluating accuracy, calibration, and efficiency together when selecting quantization strategies for clinical deployment.
APA
Ozdemir, O.B. & Li, R.. (2026). Bridging the Reliability Gap: INT8 Quantization Effects on Discrimination and Calibration in Medical Imaging. Proceedings of the 7th Conference on Health, Inference, and Learning, in Proceedings of Machine Learning Research 333:952-965 Available from https://proceedings.mlr.press/v333/ozdemir26a.html.

Related Material