RELIABILITY AND DECISION SUPPORT OF ARTIFICIAL INTELLIGENCE IN CHEST RADIOGRAPHY UNDER TECHNICAL IMAGE SHIFTS
Chest radiograph AI is susceptible to changes that have to do with the imaging process itself. This susceptibility not only affects discrimination but also calibration, coverage, and action selection. The current paper aims at studying and enhancing the decision-aware robustness contour of artificial intelligence in chest radiography involving structured technical variations. For this aim, the central internal contour was designed based on the data in the open source MIMIC dataset, whereas the major external continuation lines were developed by relying on the data in CheXpert and NIH ChestXray14 datasets. To design such a system, a set of tools has been used, which includes, among others, model-family comparison, technical shift construction, probabilistic shift detection, representation manipulation, calibration, abstention, and policy optimization with respect to expected harm. Technical shifts of various types included heel-like, scatter-like modifications, as well as secondary exposure, Poisson noise, motion blur, and stress mixture. As for the model families, convnet baseline, generic vision transformers, CLIP-like contrastive learning models, and chest radiography foundation encoders were involved. The mainline selected internally achieved improvements in AUROC from 0.816 to 0.838, AUPRC from 0.667 to 0.700, expected harm from 0.109 to 0.083, and calibration error from 0.035 to 0.021. Regarding the technical shift, the damage was not very significant when the pattern of disturbance was similar to the heel pattern as opposed to the scatter pattern, while the optimized policy that involves several actions minimized the damage to 0.101 to 0.082 in contrast to the conventional selective classifier. The practical significance of this study is that the reliability of the technical shift has been improved, not only at the level of classifiers but also at the level of decision-making through minimizing damage.
Сведения
- Журнал
- Вестник КазАТК
- Опубликовано
- 2026 · Т. 144, № 3, С. 300-312
- Язык
- Английский
- ISSN
- 1609-1817 · 2790-5802
- Издатель
- JSC Academy of Logistics and Transport