Evaluation of Grad-CAM for explaining Deep Learning's decisions on various medical imaging datasets
Session
Computer Science and Communication Engineering
Description
Deep Learning (DL) is a well-established pipeline for feature extrac- tion in medical and non-medical imaging tasks, such as object detection, segmen- tation, and classification. However, DL faces the issue of explainability, which
prohibits reliable utilisation in everyday clinical practice. The present study em- ploys the well-established Grad-CAM algorithm to assess the decisions of a Deep
Learning framework in various medical image classification tasks. Eleven da- tasets are utilised, involving images from SPECT, CT, Microscopy, and X-Ray,
which correspond to numerous diseases, including Lung Cancer, Coronary Ar- tery Disease, and COVID-19. The main conclusion of the research is that DL
with Grad-CAM might reveal important image features. However, it is observed
that on many occasions, Grad-CAM shows the model’s inefficiency in discover- ing the right locations, even in the classification accuracy is at a top level.
Keywords:
Deep Learning; Grad-CAM; Explainability
Proceedings Editor
Edmond Hajrizi
ISBN
978-9951-550-50-5
Location
UBT Kampus, Lipjan
Start Date
29-10-2022 12:00 AM
End Date
30-10-2022 12:00 AM
DOI
10.33107/ubt-ic.2022.283
Recommended Citation
Athanasoula, Ifigeneia; Apostolopoulos, Ioannis D.; and Groumpos, Peter P., "Evaluation of Grad-CAM for explaining Deep Learning's decisions on various medical imaging datasets" (2022). UBT International Conference. 292.
https://knowledgecenter.ubt-uni.net/conference/2022/all-events/292
Evaluation of Grad-CAM for explaining Deep Learning's decisions on various medical imaging datasets
UBT Kampus, Lipjan
Deep Learning (DL) is a well-established pipeline for feature extrac- tion in medical and non-medical imaging tasks, such as object detection, segmen- tation, and classification. However, DL faces the issue of explainability, which
prohibits reliable utilisation in everyday clinical practice. The present study em- ploys the well-established Grad-CAM algorithm to assess the decisions of a Deep
Learning framework in various medical image classification tasks. Eleven da- tasets are utilised, involving images from SPECT, CT, Microscopy, and X-Ray,
which correspond to numerous diseases, including Lung Cancer, Coronary Ar- tery Disease, and COVID-19. The main conclusion of the research is that DL
with Grad-CAM might reveal important image features. However, it is observed
that on many occasions, Grad-CAM shows the model’s inefficiency in discover- ing the right locations, even in the classification accuracy is at a top level.