EUVIMEDEUVIMED entdecken
Crossref

Calibration as a Search Objective in Mixed-Precision Post-Training Quantization of Medical Image Classifiers: An Empirical Multi-Objective Study

Aymen Jalil Abdulelah

Journal of Al-Farabi for Engineering Sciences · 01.12.2025 · 4 (2) · 462-470

DOI: 10.59746/3qyv8642 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Al-Farabi for Engineering Sciences, online erschienen am 01.12.2025. Verfasst von Aymen Jalil Abdulelah.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Post-training quantization allows diagnostic neural networks to run on the inexpensive processors available in many hospitals without retraining. Per-layer bit-widths are usually chosen for accuracy and computational cost only, although quantization can also change how well predicted probabilities match observed accuracy. This study makes calibration a third search objective. The selection of weight and activation bit-widths is formulated as a three-objective problem (classification error, bit operations and expected calibration error after temperature scaling) under an accuracy floor, and is solved with a reference-point-based evolutionary search whose initialisation and mutation are guided by a per-layer sensitivity table that includes calibration. The experiments cover three MedMNIST tasks (dermatoscopy, paediatric chest radiography and blood-cell microscopy) and two networks (ResNet-18 and MobileNetV2). The search is compared over ten seeds with uniform precision, greedy sensitivity allocation, a Hessian-based integer program, random search and the same search without the calibration objective, and robustness is tested on corrupted test images. The whole study ran on a four-core laptop processor in about 17.7 hours. The proposed search found fronts with a higher hypervolume than random search, uniform precision and the integer program in all six dataset-network combinations, while greedy allocation had the best average rank. The selected models used about 2% of the full-precision bit operations and lost at most 0.017 in the area under the receiver operating characteristic curve, and removing the sensitivity guidance lowered the hypervolume by 23% to 30%. The calibration objective reduced the validation calibration error in every combination, but the test calibration error in only four of six, with one statistically significant difference. Under image corruptions the gains were consistent in direction but small. Calibration should therefore be checked after quantization rather than assumed.

Zitieren

APA

Abdulelah, A. J. (2025). Calibration as a Search Objective in Mixed-Precision Post-Training Quantization of Medical Image Classifiers: An Empirical Multi-Objective Study. Journal of Al-Farabi for Engineering Sciences, 4(2), 462-470. https://doi.org/10.59746/3qyv8642

Vancouver

Abdulelah AJ. Calibration as a Search Objective in Mixed-Precision Post-Training Quantization of Medical Image Classifiers: An Empirical Multi-Objective Study. Journal of Al-Farabi for Engineering Sciences. 2025;4(2):462-470. doi:10.59746/3qyv8642

BibTeX
@article{Abdulelah2025,
  author = {Abdulelah, Aymen Jalil},
  title = {Calibration as a Search Objective in Mixed-Precision Post-Training Quantization of Medical Image Classifiers: An Empirical Multi-Objective Study},
  journal = {Journal of Al-Farabi for Engineering Sciences},
  year = {2025},
  volume = {4},
  number = {2},
  pages = {462--470},
  doi = {10.59746/3qyv8642}
}
Kostenlos starten →