A Comparative Deep Learning Framework for Brain Tumour Detection and Sub-Region Segmentation in MRI
Dr. Mungara Kiran Kumar,
B Lakshmi Nivas,
D Jaswant,
K Himabindu,
Shubh Nigam
This study introduces a step-by-step upgrade path for segmenting three brain tumor classes on BraTS
2023 GLI magnetic resonance images, an expansion of the standard BraTS reference set. The first version
is a plain U-Net. The pipeline then adds one upgrade at a time - a deeper encoder to extract richer
features, a hybrid Dice-cross-entropy loss to counter class skew and attention gates to sharpen spatial
focus. Every stage is tested under an identical protocol so that the gain from each change is measured in
isolation. After all stages, the model records a higher Dice score and a higher Intersection over Union
value. The final Attention U-Net with hybrid loss draws crisper borders and finds the smallest tumor
zones more often, which supports the value of the sequential refinement plan.
