Enhancing Deep Learning-Based Subabdominal MR Image Segmentation During Rectal Cancer Treatment: Exploiting...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• U-Net models face challenges in subabdominal MR image segmentation due to misalignment and semantic gaps from convolutional and pooling operations.
• These issues are particularly relevant in rectal cancer treatment, impacting diagnostic accuracy.

Purpose of the Study:

• To develop an improved MR image segmentation method for rectal cancer treatment.
• To address and mitigate misalignment and semantic gap problems inherent in U-Net architectures.

Main Methods:

• Proposed a novel MR image segmentation approach utilizing a multiscale feature pyramid network and a bidirectional cross-attention mechanism.
• Incorporated dilated convolution and a multiscale feature pyramid network in the encoding phase to reduce the semantic gap.
• Implemented a bidirectional cross-attention mechanism to preserve spatial information and minimize misalignment within the U-Net framework.

Main Results:

• The proposed method demonstrated superior performance compared to existing techniques on a subabdominal MR image dataset.
• Experimental results validated the effectiveness of the novel approach in enhancing segmentation accuracy.

Conclusions:

• A multiscale feature pyramid network effectively reduces the semantic gap in MR image segmentation.
• The bidirectional cross-attention mechanism facilitates improved feature alignment between encoding and decoding stages, enhancing overall segmentation quality.