Prostate lesion segmentation based on a 3D end-to-end convolution neural network with deep multi-scale attention.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer diagnosis relies heavily on accurate lesion segmentation from multi-parametric Magnetic Resonance Imaging (mp-MRI).
• Current segmentation methods are often multi-step, leading to time-consuming processes, error accumulation, and reduced accuracy.
• Small lesion size, irregular shapes, and blurred contours present significant challenges in prostate lesion segmentation.

Purpose of the Study:

• To develop an efficient and accurate automatic prostate lesion segmentation method.
• To reduce computation time and improve the fusion of multi-level contextual information from mp-MRI.
• To introduce an end-to-end deep learning approach for prostate cancer segmentation.

Main Methods:

• Proposed a novel end-to-end deep learning framework, DMSA-V-Net, integrating all segmentation steps into a single process.
• Utilized a 3D V-Net as the backbone architecture, representing the first application of 3D CNNs for this task.
• Incorporated a deep multi-scale attention mechanism to enhance focus on the region of interest (ROI) and suppress background noise.

Main Results:

• Achieved a Dice score of 0.7014 and a sensitivity of 0.8652 in experiments involving 97 patients across five cross-fold validations.
• The DMSA-V-Net demonstrated superior performance compared to existing segmentation methods.
• The attention mechanism adaptively realigned contextual information across different feature map scales and high-level saliency maps.

Conclusions:

• The proposed DMSA-V-Net offers a significant advancement in automatic prostate lesion segmentation from mp-MRI.
• The end-to-end approach effectively reduces computation time and enhances segmentation accuracy.
• This method holds promise for improving the diagnosis and management of prostate cancer.