Explainability in Computer Vision Comparing Grad-CAM++ to CAM, Guided Grad-CAM and Integrated Gradients on the explainability of image classification on the Mini ImageNet dataset and the Indian Diabetic Retinopathy Image Dataset

Dortant, B.

Explainability in Computer Vision Comparing Grad-CAM++ to CAM, Guided Grad-CAM and Integrated Gradients on the explainability of image classification on the Mini ImageNet dataset and the Indian Diabetic Retinopathy Image Dataset

Files

Dortant, B. s-4484754-BSc-Thesis-2024.pdf (6.81 MB)

Authors

Dortant, B.

Issue Date

2024-08-14

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/20707

Abstract

Abstract Despite the unparalleled success of convolutional neural networks in computer vision tasks, their lack of interpretability prevents their use in the medical domain. Several methods exist to visualize the important parts of the image for the neural decision making, and thereby enhancing the explainability of the arti cial neural network predictions. In this thesis we will test and compare the performance of several explainability methods: CAM, Grad-CAM++, Integrated Gradients, and Guided Grad-CAM. For the evaluation both Mini ImageNet [24] and the indian diabetic retinopathy image dataset [17] on kaggle were used. For the Mini ImageNet dataset we see opposing values for di erent metrics, which generalized from training to validation and test set. Therefore, there is not a clear best method to explain a decision made by a convolutional neural network on image data. The same holds for the indian diabetic retinopathy image dataset. The results did roughly generalize from training to validation and test set, but the results did not generalize between the two datasets. This could be due to the small size of the indian diabetic retinopathy image dataset.