Machine Learning and Deep Learning for Brain Tumor Prediction and Segmentation in MRI: A Systematic Review of Methods, Datasets and Open Challenges

Magnetic resonance imaging (MRI) is the reference modality for brain tumor diagnosis, yet manual delineation remains slow, expertise-dependent and subject to inter-observer variability. Over the last four years the literature has shifted decisively from handcrafted feature pipelines to convolutional, encoder-decoder, transformer and hybrid learning architectures. This paper presents a systematic review of forty-one primary studies published between 2023 and 2026, selected from 412 records through a PRISMA-style protocol. The reviewed work is organised into a four-branch taxonomy - classical machine learning, CNN backbones, encoder-decoder segmentation networks and transformer/hybrid ensembles - and is compared on a common footing of accuracy, precision, recall, F1-score, Dice similarity coefficient and intersection over union. Reported classification accuracy ranges from 93.30% to 99.70% and Dice from 0.650 to 0.961, but the spread is shown to be driven as much by dataset choice and evaluation protocol as by architecture. A reference hybrid detection-segmentation pipeline that couples classical preprocessing and region-based localisation with CNN feature learning and a margin-based or meta-learned decision layer is formalised as two algorithms, and priorities for trustworthy, federated and explainable neuro-oncology AI are set out.