The week's pick
Random Articles
Reseach Article
Does Probability Calibration Survive a Change of OCT Scanner? A Leakage-Free Study of Calibration Transfer Across Manufacturers
| International Journal of Computer Applications |
| Foundation of Computer Science (FCS), NY, USA |
| Volume 187 - Number 149 |
| Year of Publication: 2026 |
| Authors: Tanha Asfika Jaman, Akter Mst Lovely, Iftee Shekh Iftesham, Mostafa Farzana |
10.5120/ijca8fee57096b16
|
Tanha Asfika Jaman, Akter Mst Lovely, Iftee Shekh Iftesham, Mostafa Farzana . Does Probability Calibration Survive a Change of OCT Scanner? A Leakage-Free Study of Calibration Transfer Across Manufacturers. International Journal of Computer Applications. 187, 149 ( Oct 2026), 37-44. DOI=10.5120/ijca8fee57096b16
Abstract
Retinal optical coherence tomography (OCT) classifiers are usually reported with calibration measured on the same scanner on which they were trained. This study measures what happens to their predicted probabilities on a different manufacturer’s scanner. Three-class classifiers (AMD, DME, normal) are trained on 65,585 patient-disjoint Heidelberg Spectralis B-scans (Kermany) and evaluated on an unseen Optovue Avanti RTVue XR cohort (OCTDL) across eleven runs, two backbones and two resizing policies. The data audit finds two leakage channels: Kermany patient identifiers are global, being shared between class folders 501 to 2,597 times more often than chance, and in OCTDL 47 pairs of pixel-identical images carry different patient identifiers. Splits are therefore built on connected components of (same patient) OR (near-duplicate). The resizing policy matters most for accuracy: square resizing raises target balanced accuracy from 0.8821 to 0.9304 and lowers uncalibrated expected calibration error (ECE) from 0.0316 to 0.0181 relative to letterboxing, a reduction comparable to that of the best source-only calibrator. Source-only temperature scaling lowers target ECE in 10 of 11 runs (95% bootstrap interval excluding zero in 6); calibrating on source data degraded by a physically parameterised OCT acquisition model beats a generic corruption family of matched severity in only 5 of 11 runs; and input-conditioned temperature scaling loses to a single global temperature in 10 of 11 runs. A referral threshold promising 0.5% error on source data yields 1.19% to 1.59% error on the new scanner.
References
- Vijayan, M., Veerapu, G., and Nisha, J. S. 2026. A spatial adaptive multi-scale ConvNeXt framework for robust, calibrated, and statistically validated multi-class retinal OCT classification. Frontiers in Artificial Intelligence, 9. DOI: 10.3389/frai.2026.1898978.
- Kumar, A. 2026. Calibrated hybrid CNN-Transformer for retinal OCT classification. arXiv:2607.09809.
- Chutani, D. and Yalavarthy, P. K. 2026. Noise-aware dynamic convolution for improved generalizability of retinal disease diagnosis using optical coherence tomography images. Journal of Biomedical Optics, 31(8), 086006. DOI: 10.1117/1.JBO.31.8.086006.
- Lopukhova, E. A., Idrisova, G. M., Mukhamadeev, T. R., Voronkov, G. S., Kutluyarov, R. V., and Topolskaya, E. P. 2026. A hierarchical deep learning architecture for diagnosing retinal diseases using cross-modal OCT to fundus translation in the lack of paired data. Journal of Imaging, 12(1), 36. DOI: 10.3390/jimaging12010036.
- Morano, J., Fazekas, B., et al. 2025. Multimodal foundation model and benchmark for comprehensive retinal OCT image analysis. npj Digital Medicine, 8, 576. DOI: 10.1038/s41746-025-01852-3.
- Salvador, T., Voleti, V., Iannantuono, A., and Oberman, A. 2021. Frustratingly easy uncertainty estimation for distribution shift. arXiv:2106.03762.
- Zhang, Y., Xu, C., Wu, Y., Guan, Z., and Zhao, W. 2025. Target-agnostic calibration under distribution shift with frequency-aware gradient rectification. arXiv:2508.19830.
- Hendrycks, D. and Dietterich, T. 2019. Benchmarking neural network robustness to common corruptions and perturbations. In Proc. ICLR. arXiv:1903.12261.
- Agnihotri, S., Schader, D., Sharei, N., Kacar, M. E., and Keuper, M. 2025. Are synthetic corruptions a reliable proxy for real-world corruptions? In CVPR 2025 Workshop on Synthetic Data for Computer Vision. arXiv:2505.04835.
- Tampu, I. E., Eklund, A., and Haj-Hosseini, N. 2022. Inflation of test accuracy due to data leakage in deep learning-based classification of OCT images. Scientific Data, 9, 580. DOI: 10.1038/s41597-022-01618-6.
- Guo, C., Pleiss, G., Sun, Y., and Weinberger, K. Q. 2017. On calibration of modern neural networks. In Proc. ICML. arXiv:1706.04599.
- Ovadia, Y., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J. V., Lakshminarayanan, B., and Snoek, J. 2019. Can you trust your model's uncertainty? Evaluating predictive uncertainty under dataset shift. In Proc. NeurIPS. arXiv:1906.02530.
- Wang, X., Long, M., Wang, J., and Jordan, M. I. 2020. Transferable calibration with lower bias and variance in domain adaptation. In Proc. NeurIPS. arXiv:2007.08259.
- Ding, Z., Han, X., Liu, P., and Niethammer, M. 2021. Local temperature scaling for probability calibration. In Proc. ICCV. arXiv:2008.05105.
- Joy, T., Pinto, F., Lim, S.-N., Torr, P. H. S., and Dokania, P. K. 2023. Sample-dependent adaptive temperature scaling for improved calibration. In Proc. AAAI. arXiv:2207.06211.
- Khan, B. and Syed, T. Q. 2025. Confidence-calibrating regularization for robust brain MRI segmentation under domain shift. arXiv:2509.23176.
- Di Salvo, F., Doerrich, S., and Ledig, C. 2024. MedMNIST-C: Comprehensive benchmark and improved classifier robustness by simulating realistic image corruptions. arXiv:2406.17536.
- Kermany, D. S., Goldbaum, M., Cai, W., et al. 2018. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5), 1122–1131. DOI: 10.1016/j.cell.2018.02.010.
- Kulyabin, M., Zhdanov, A., Nikiforova, A., et al. 2024. OCTDL: Optical coherence tomography dataset for image-based deep learning methods. Scientific Data, 11, 365. DOI: 10.1038/s41597-024-03182-7.
Index Terms
Keywords