JMCD Dataset for Brain Tumor Detection and Analysis Using Explainable Deep Learning

The growth of abnormal cells in the brain gives rise to a deadly form of cancer known as a brain tumor. The mass of brain tumors proliferates and rises very fast, and if not appropriately treated, the patient’s survival rate is less or can rapidly lead to death. A very exigent task for radiologists is early brain tumor detection which may help to evaluate the tumor and plan treatment for an effective prognosis. In this work, we used deep learning to detect and localize brain tumors from MRI slices. After the detection of the tumor, we generate a color heatmap to show the tumor’s location, severity, and possible affected regions. Our novel image processing algorithm swiftly segments the skull region from the background of MR images, augmented over heatmap for better observation. We analyzed the performance of VGG 16, Resnet 50, and EfficientNet deep learning architectures. Transfer learning on Resnet 50 model with custom top layers achieved excellent performance with 99.37% accuracy, 99.68% precision, and 99.69% specificity, which is better than most of the works in brain tumor detection using machine learning. Expert radiologists critically examined our model outcomes in detail. They found our method effective to triage people with brain tumors and helpful in decision making for the clinicians by severity indication through heatmaps. They observed the prime cause for false positives to be unbalanced contrast in some MR images. Through this work, we also made the novel dataset named as JMCD (Jabalpur Medical College Dataset) publicly available for research which contains all four MR pulse sequences annotated by expert radiologists for 140 patients. Each patient has 15–20 sequences containing 250 images. Through this work and accompanying data, we aim to give better methods for brain tumor detection and help research in the same field. © 2023, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.