Our primary focus is on developing well-calibrated out-of-distribution (OOD) detectors to ensure the safe deployment of medical image classifiers. The use of synthetic augmentations has become common for specifying regimes of data inliers and outliers. However, our research findings highlight the substantial influence of both the synthesis space and the type of augmentation on the performance of OOD detectors. After conducting an extensive study using medical imaging benchmarks and open-set recognition settings, we recommend employing a combination of virtual inliers in the classifier's latent space and diverse synthetic outliers in the pixel space. This approach proves highly effective in producing OOD detectors with superior performance.
Our calibration protocol consistently outperforms all baselines (G-ODIN, VOS, VOS++, NDA, NDA++) by significant margins (10% to 30% on average for Modality Shift Detection and 15% to 28% for Novel Class Detection ) in terms of AUROC scores without compromising on the ID accuracy.
@inproceedings{narayanaswamy2023know,
title={Know Your Space: Inlier and Outlier Construction for Calibrating Medical {OOD} Detectors},
author={Vivek Narayanaswamy and Yamen Mubarka and Rushil Anirudh and Deepta Rajan and Andreas Spanias and Jayaraman J. Thiagarajan},
booktitle={Medical Imaging with Deep Learning},
year={2023}}
If you have any questions, please feel free to contact us via email: narayanaswam1@llnl.gov; jjayaram@llnl.gov