•  
  •  
 

Subject Area

Computer Science

Article Type

Original Study

Abstract

The critical need for accurate and robust brain tumor detection in MRI video scans is for early diagnosis and treatment planning operations. However, the existing methods suffer from several limitations, such as not being able to capture spatio-temporal features, handling class imbalances, and generalizing across datasets with diverse characteristics. To address these challenges, I propose a hybrid deep learning framework that integrates multiple advanced techniques to enhance the detection and segmentation of brain tumors. My approach starts by using the 3D CNN-LSTM (ResNet3D-LSTM) to extract spatio-temporal feature, which utilizes ResNet3D for spatial detail extraction and LSTM for temporal coherence across MRI frames; this avoids a lack of proper modeling of MRI video data samples' sequential structure. The use of an attention-based key frame detection module with the Transformer Encoder reduces the computational overhead by 30%-40% and highlights interpretability with the use of attention heatmaps for focusing attention on diagnostically significant frames. For accurate tumor segmentation, I propose the integration of Feature Pyramid Networks with UNet for multi-scale feature fusion that captures both fine-grained and global spatial features. The presented framework here achieves state-of-the-art results by realizing classification accuracies of 95% and 97% and having Dice similarity coefficients ranging from 0.89 to 0.93, enhancing cross- domain generalization. This model proposed sets a new benchmark for MRI video-based brain tumor detection, significantly advancing diagnostic accuracy, computational efficiency, and clinical applicability sets.

Keywords

Brain Tumor Detection, Spatio- Temporal Features, 3D CNN-LSTM, Attention Mechanism, Domain Adaptations

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS