Abstract
CT X-ray tubes are among the most expensive consumables in medical imaging, and in low-resource hospitals each unplanned failure causes weeks of scanner downtime. Predictive maintenance through Remaining Useful Life (RUL) estimation could prevent this, but real hospital logs record only sparse failure events, not the continuous sensor data deep models need. We calibrate a physics-based degradation simulator to real failure records from a Moroccan public hospital (CHU Ibn Rochd), reproducing the observed mean tube lifespan within 4.0%, and use it to generate synthetic RUL-labelled trajectories. On these we benchmark six models, from classical regressors to a physics-informed LSTM (PINN-LSTM), under progressively scarcer training data (100%, 50%, 20%) over five seeds. Two findings emerge. First, model simplicity drives robustness: as data shrinks to 20%, lightweight recurrent models stay below 3.8 days RMSE while CNN-LSTM and Transformer architectures exceed 5.4 days. Second, the physics-informed monotonicity constraint matches the plain LSTM’s accuracy and improves physical validity only when data is abundant, offering no advantage under scarcity; a C-MAPSS replication confirms its effect is small and dataset-dependent. In data-scarce clinical settings, simpler models are more reliable than added architectural or physics-based complexity.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_010.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=xPzjeyyEZz
BibTex
@InProceedings{OubZak_CT_MICCAISAT2026,
author = { Oubnini, Zakaria AND Wang, Feng},
title = { { CT X-ray Tube RUL Estimation Under Data Scarcity in Low-Resource Hospitals } },
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}
}