Tumorxnet++-Roi: Region-Of-Interest–Driven Deep Learning Framework For Tumor Size Estimation And Survival Prediction In Brain Mri

Accurate​‍​‌‍​‍‌ measurement of the extent of a brain tumour and estimation of survival time is an essential part of neuro-oncological imaging. This research, which is built on our previous TumourXNet++ architecture for tumour detection, presents TumorXNet++-ROI, a region-of-interest (ROI) guided deep learning system that combines tumour segmentation, morphometric feature extraction, and prognostic modeling in a single framework. Initially, the proposed method of isolating tumour ROIs by adaptive segmentation is followed by the extraction of quantitative descriptors-tumour area, maximal width, height, and mean density-that act as interpretable surrogates of lesion growth. A neural regression module is used to estimate tumour size, while a second survival-prediction head makes an estimate of the post-diagnostic lifespan directly from the ROI features and shallow CNN embeddings. The experiments involved 501 BraTS-2021 MRI studies to compare the estimated and model-predicted survival values. There were strong inverse correlations between tumour size indicators and survival (Pearson’s r: area = −0.975, width = −0.908, height = −0.908, density = −0.686) whereby bigger and denser tumours led to lower predicted survival. The regression analysis produced a high coefficient of determination (R² > 0.90) with a low RMSE and MAE, and the Bland–Altman assessment indicated tight agreement intervals between the predicted and the reference survival. These results indicate that a deep learning method based on ROI normalization makes it possible to accurately quantify a tumour and to estimate survival time based on routine MRI data. The modular TumourXNet++-ROI pipeline thus moves beyond classification-focused AI systems to clinically interpretable prognosis modelling, providing a reproducible basis for the development of 3D volumetric and multimodal survival-analysis frameworks in the ​‍​‌‍​‍‌future.