Attention-enabled 3D boosted convolutional neural networks for semantic CT segmentation using deep supervision.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy Planning

Background:

• Accurate semantic segmentation of anatomical structures in CT images is crucial for effective radiotherapy planning.
• Current state-of-the-art convolutional neural networks (CNNs) face challenges in prioritizing relevant anatomy and managing network complexity for segmentation tasks.

Purpose of the Study:

• To introduce a novel Deeply Supervised Attention-enabled Boosted Convolutional Neural Network (DAB-CNN) for superior semantic CT segmentation.
• To evaluate the performance of DAB-CNN against seven other leading CNN architectures in segmenting prostate, rectum, and penile bulb.

Main Methods:

• Development of a 3D cascaded CNN framework incorporating spatial attention gates (AGs) and incremental channel boosting.
• Implementation of deep supervision to enhance feature learning and model convergence.
• Comparative analysis of DAB-CNN with U-Net variations, A-UNet, B-UNet, D-UNet, ResNeXt, LL-CNN, and DA-UNet using Dice scores on a dataset of 120 prostate radiotherapy patients.

Main Results:

• DAB-CNN achieved significantly superior Dice scores compared to all alternative algorithms for segmenting the prostate, rectum, and penile bulb.
• The proposed model demonstrated enhanced accuracy in prioritizing relevant anatomy and suppressing redundancies through attention mechanisms.

Conclusions:

• Attention-enabled boosted CNNs with deep supervision, as exemplified by DAB-CNN, represent a superior approach for automatic semantic CT segmentation.
• DAB-CNN offers improved prediction accuracy over current state-of-the-art methods, benefiting radiotherapy planning.