Back

Uncertainty Quantification in COVID-19 Detection Using Evidential Deep Learning

Khosravi, B.; Faghani, S.; Ashraf Ganjouei, A.

2022-05-29 radiology and imaging
10.1101/2022.05.29.22275732 medRxiv
Show abstract

Considering the immense pace of developments in deep learning (DL), its applications in medicine are relatively limited. One main issue that hinders the utilization of DL in the medical practice workflow is its reliability. A radiologist interpreting an image can easily say "I dont know", while a DL model is forced to output a result. Evidential deep learning (EDL) is one of the methods for uncertainty quantification (UQ). In this work, we aimed to use EDL to express model uncertainty in detecting COVID-19. We used SIIM-FISABIO-RSNA COVID-19 chest x-ray dataset and trained a model to diagnose typical COVID-19 pneumonia. When applied to a separate test set, it yielded an accuracy of 88% with median uncertainty scores of 0.25 and 0.07 for normal and typical COVID-19 images, respectively. Moreover, the model labeled unseen indeterminate and atypical COVID-19 x-rays with median uncertainties of 0.32 and 0.35, respectively. Our models performance was superior to the exact model trained with conventional approach of DL (i.e., using the cross-entropy loss), which is not able to express the uncertainty level. Overall, this study demonstrates applicability of UQ in disease detection that could facilitate the use of DL in practice by increasing its reliability.

Matching journals

The top 8 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.