Recalibrating 3D ConvNets with Project & Excite.

Area of Science:

• Computer Vision
• Medical Imaging
• Deep Learning

Background:

• Fully Convolutional Neural Networks (F-CNNs) are state-of-the-art for image segmentation.
• Squeeze and Excitation (SE) blocks enhance F-CNNs by recalibrating feature maps.
• Existing SE blocks are primarily designed for 2D architectures, limiting their application to volumetric medical data.

Purpose of the Study:

• To extend 2D recalibration methods to 3D for volumetric medical image segmentation.
• To introduce novel Project & Excite (PE) modules tailored for 3D F-CNNs.
• To evaluate the performance of PE modules against existing recalibration techniques in 3D F-CNNs.

Main Methods:

• Developed a generic compress-process-recalibrate pipeline for comparing 3D recalibration blocks.
• Introduced Project & Excite (PE) modules that compress feature maps along spatial dimensions, retaining more information.
• Integrated PE modules into 3D F-CNNs for segmentation tasks.

Main Results:

• PE modules significantly boosted segmentation performance, achieving up to a 0.3 increase in Dice Score.
• PE modules outperformed 3D extensions of other recalibration blocks.
• The integration of PE modules resulted in only a marginal increase in model complexity.

Conclusions:

• Project & Excite modules offer an effective way to enhance 3D F-CNNs for medical image segmentation.
• PE modules provide superior performance compared to existing 3D recalibration methods.
• The proposed PE modules are easily integrable and minimally impact computational cost.