|
|
 |
| |
|
|
|
Enhanced Brain Tumor Segmentation Using Multi- Stream Hybrid Deep Learning with Cross-Attention Integration |
|
|
|
PP: 619-627 |
|
|
|
Author(s) |
|
|
|
Sulieman Shelash Mohammad,
Hamza Abu Owida,
Suhaila Abuowaida,
Asokan Vasudevan,
Nawaf Alshdaifat,
Mohammad Faleh Ahmmad Hunitie,
|
|
|
|
Abstract |
|
|
| This paper proposes a novel hybrid deep learning architecture, the Multi-Stream Cross Attention Network (MSCAN), for brain tumor segmentation from multi-modal MRI images. MSCAN employs a multi-stream feature extraction branch that processes multiple MRI modalities in parallel and then combines them with a cross-attention mechanism. This enables the model to dynamically weight significant features across modalities, enhancing segmentation accuracy. Besides, MSCAN enhances computational efficiency by sharing representations across streams and avoiding redundancy. Evaluated on the BraTS 2021 dataset, MSCAN achieves Dice similarity coefficients of 0.92 for whole tumors, 0.89 for tumor cores, and 0.85 for enhancing tumors, outperforming the state of the art by over 3.5%. Moreover, MSCAN reduces computational overhead by 28% with high segmentation accuracy. These outcomes verify MSCAN’s efficacy for both accuracy and efficiency improvement of clinical brain tumor segmentation. |
|
|
|
|
|
|
|
