A Multi-Modal Hybrid U-Net and Mask R-CNN Approach to Brain Tumor Segmentation in MRI: Baseline Analysis and Proposed Framework
DOI:
https://doi.org/10.70917/ijcisim-2026-5108Keywords:
Brain tumor segmentation, Multi-modal MRI, U-Net, Mask R-CNN, BraTS, Deep learning, Dice coefficient, False-positive reduction, Dynamic thresholdingAbstract
Detecting brain tumors early is still hard to get right, and it matters a lot because it feeds directly into how a patient gets treated. MRI is the modality most people trust for this because it shows soft tissue clearly, but drawing tumor boundaries by hand takes forever, and two radiologists rarely agree exactly. That disagreement problem is a big part of why automated segmentation has become such an active research area recently. In this paper, I look at where current deep learning methods for MRI tumor segmentation are still weak: they usually depend on just one MRI sequence, false positives don't go away easily, tumor boundaries are often imprecise (HD95 numbers make this obvious), and diffusion-based models in particular cost too much compute to run quickly. I implemented two baseline pipelines on BraTS 2020 to check where things actually stand - a GGMM-enhanced U-Net and a diffusion probabilistic model - and used what came out of that to shape a hybrid framework. The framework has four parts: it fuses T1, T2, FLAIR, and T1CE features instead of using just one sequence; it uses a Dice-BCE loss so class imbalance doesn't dominate training; it applies a threshold that adjusts itself based on the mean and standard deviation of each prediction rather than sticking to a fixed 0.5 cutoff; and it removes tiny, implausible regions in a post-processing step. The GGMM baseline landed around a 0.81-0.82 Dice score; the diffusion model got closer to 0.89 but took far longer per case to run. Based on that gap, four hypotheses are proposed for testing once the full model is trained: Dice of 0.90 or higher, IoU above 0.85, at least 15% fewer false positives than a plain U-Net, and no more than a 5% performance drop under noisy input. The hope is that this ends up being a reasonable middle ground - accuracy near what diffusion models get, minus the runtime problem that keeps them out of actual clinics.
