Research Highlights

Attention-reﬁned U-Net with Skip Connections for Eﬀective Brain Tumor Segmentation from MRI Images

Authors: A. F. M. Minhazur Rahman and Md. Ali Hossain

Figure 1: SC-SE U-Net architecture.

Contributions:

Developed SC-SE U-Net, integrating Squeeze-and-Excitation channel attention blocks and skip connections similar to residual network
Improved segmentation accuracy with an Intersection over Union (loU) score of 84.22%

Authors: A. F. M. Minhazur Rahman and Md. Ali Hossain

Contributions:

Experimented with fine-tuned EfficientNet, DenseNet, VGG, and ResNet encoder architectures in a modified U-Net architecture
Combined outputs from top models through majority voting and weighted voting ensembles for robust segmentation results
Achieved an loU score of 85.15%

Authors: A. F. M. Minhazur Rahman and Boshir Ahmed

Contributions:

Introduced the FA-CNN method, combining factor analysis for dimension reduction with CNNs for spectral-spatial feature extraction
Achieved high classification accuracy with 99.59% overall accuracy on the Indian Pines dataset and 99.95% on the Pavia University dataset

Authors: Kowshik Deb Nath, A. F. M. Minhazur Rahman and Md. Ali Hossain

Figure 2: Attention-based CNN architecture for knee injury classiﬁcation.

Contributions:

Introduced an attention-based CNN model achieving 100% accuracy in binary classification and 91% in multi-class classification
Demonstrated model effectiveness on two comprehensive knee MRI datasets

Project Title: Brain Tumor Segmentation Using Effective Deep Learning Approach

Financed By: University Grants Communication, Bangladesh and Rajshahi University of Engineering & Technology

Duration: 2022-2023

Figure 3: EﬃcientNet-SE U-Net architecture.

Contribution:

Proposed a novel EfficientNet-SE U-Net model combining EfficientNetB4 encoder and Squeeze-and-Excitation (SE) blocks in decoder
Achieved high performance with loU of 85.34%
Enhanced feature extraction using transfer learning
Improved segmentation precision with SE blocks