XAIMed-Net: towards explainable brain tumour detection in 2D T1-weighted CE-MRI images using transfer learning

Malignant brain tumours comprise one of the deadliest types of cancer and their timely and accurate detection is receiving urgent attention. Radiology can benefit from the recent impressive progress of AI and deep learning that provide the non-invasive diagnostic opportunity. However, the challenge of explainable AI-based decision-making processes remains. In this work, we aim to take a step toward more explainable classification of brain tumours in medical images by designing a new convolutional architecture while also leveraging the advantages of transfer learning. Specifically, we introduce a new shallow convolutional head that can be added to a pre-trained ResNet50 model to produce EigenCAM-based importance heatmaps. We investigate pre-training based on the ImageNet and RadImageNet datasets. Furthermore, we also address the problem of weakly-supervised brain tumour segmentation in a challenging 2D T1-weighted CE-MRI dataset by developing an approach to post-process the resulting heatmaps. Our preliminary experimental results with the Dice Score Coefficient of 0.145 and the new objective distance metricµ▵ = 59.55 show the potential of combining large pre-trained models with shallow convolutional architectures to localize tumours using only weak classification labels. Our code is publicly available at https://github.com/juliadietlmeier/Explainable_brain_tumor_detection_in_MRI.