On the application of explainable artificial intelligence based interpretability and trustworthy metrics assessment for diabetic retinopathy prediction models on large and small retinal images dataset

Background In recent years, diabetic retinopathy detection gains its importance due to its crucial role in loss of vision of people in all ages due to their life style and other physiological reasons. To address this, several researchers suggested the application of Machine Learning (ML) and Deep Learning (DL) models. Even though higher accuracy was shown by these models, the trustworthiness of the prediction results are still questionable as these decisions are directly related to the vision based experiments. Methods To evaluate the interpretability and trustworthiness of the predictions, this research work has proposed an efficient and novel framework that integrated the application of Explainable AI (XAI) with Believability, Explainability, Transparency and Actionability (BETA) metrics on the prediction models. In order to evaluate, this research work has used the large retinal images dataset namely APTOS 2019 blindness detection dataset (Karthik & Sohier, 2019) with Non-Commercial and Academic Research license ( https://www.kaggle.com/competitions/aptos2019-blindness-detection/rules#7-competition-data ) that is of 10.22 GB size. In addition, it evaluated the small sized real-time datasets from the DRIVE (100 images), CHASE_DB1 (28 images), and STARE (400 images) databases. As they are very small, in order to work on them, four additional datasets were generated for training purpose using image processing techniques such as histogram equalization, Sobel filter, Canny filter, and enhanced multispectral color composite. Image augmentation expanded the dataset from 91 to 1,000 images, with each class containing 250 images. The proposed framework has two phases. Phase I deals with the application of three most advanced deep learning models namely Visual Geometry Group 16 (VGG16), Residual Network 50 (ResNet-50), InceptionV3 and EfficientNet-B0 on the large and small datasets and find out the accuracy of those models. Phase II deals with the application of XAI Methods such as Grad-CAM, SmoothGrad and Saliency Maps to provide interpretation of the results derived from the selected best DL models. In this research, the XAI methods applied on EfficientNet-B0 for large dataset and InceptionV3 model for small dataset as they exhibited higher accuracy. Finally, this work evaluated the effectiveness of the XAI models visualization with BETA metrics such as fidelity, unambiguity and interpretability size to ensure the interpretability and trustworthiness of the models. Results The results derived from the prediction models indicated that EfficientNet-B0 has the higher accuracy. Based on the BETA metrics it is observed that Grad-CAM on EfficientNet-B0 has the highest fidelity score of 95.02%, unambiguity score of 79.52% and with interpretability size of 78.02% for the large dataset and the same Grad-CAM on InceptionV3 model has the highest fidelity score of 90.05%, unambiguity score of 65.14% and with interpretability size of 75.32% for the small dataset.