Brain Tumor Analysis and Prognosis

The explosion of generative AI technologies means that the risk of deep fake alterations to medical images - via synthetic alterations for example by adding or removing pathological areas (for example adding tumors) - has increased exponentially. Such attacks can confuse or mislead clinicians resulting in misdiagnosis as well as undermining the reliability of research and potentially being detrimental to patient safety. His research presents an intelligent framework for identifying and localizing manipulated regions in brain MRI scans using the Mask R-CNN architecture. The proposed approach aims to detect deep-fake alterations that may lead to incorrect medical interpretation and diagnostic errors. MRI images will be classified into one of four categories: real tumor; no tumor; deep Fake tumour added; and deep fake tumour removed. For both training and evaluation, we used a carefully curated dataset of approximately 1,300 brain MRIs, containing both real and synthetically manipulated examples. The proposed framework uses deep convolutional feature extraction making region proposals and performing instance segmentation to identify subtle evidence of tampering that could otherwise be imperceptible visually. An easy-to-use web-based interface created with a Next.js frontend and a python-flask backend was built to allow for easy and efficient real-time upload, analysis, and visualization of prediction masks/application of results. The experimental validation of the proposed technique shows good performance in distinguishing real from manipulated MRIs, and highlighting potential regions of false-positive or false-negative results will improve the overall trust of users in AI-based medical diagnosis. Directions for future work include expansion of dataset diversity within the source population, optimization of the speed of real-time analysis, and expanding potential applications.