MRI-Based Brain Tumor Analysis with Integrated Classification, Region Estimation, and Explainable Visualization

Deep learning-based automated analysis of brain magnetic resonance imaging (MRI) can aid analysis of brain tumors.However, systems must be interpretable, estimate regions of interest, and interact with users.This paper describes a web-based framework that includes validation of MRI, MobileNetV2-based deep learning classification, OpenCV-based region of interest estimation, Grad-CAM-based explainable visualisation of results, pseudo-3D intensity visualisation, and interactive boundary correction.The system can classify MRI images between four labels: glioma, meningioma, pituitary tumour or non-tumour.Once classified, the estimated region of interest, relative size and visual interpretations of the results are provided, along with downloadable reports.The system achieved high classification accuracy of 91.43% and macro F1-score of 91.20% when tested against 1144 images in a validation set.The average inference time on the CPU was found to be approximately 85ms for each image, from a total of 49 images.The validation step had a false rejection rate of 31.0%for valid MRI images, illustrating that better input validation is required.Due to no ground truth segmentation masks being available, the accuracy of segmentation could not be quantified.This paper demonstrates how a classification system can be created with explainable visualisations and interactive segmentation.Further testing is required with independent data, annotated segmentation masks and clinical validation before it can be considered for use clinically.