Learning normalized inputs for iterative estimation in medical image segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Existing methods often struggle with diverse image modalities and complex anatomical structures.
• Convolutional Neural Networks (CNNs) and Residual Networks (ResNets) have shown promise but require effective integration for segmentation tasks.

Purpose of the Study:

• To introduce a versatile and powerful pipeline for medical image segmentation.
• To leverage advances in CNNs and ResNets for improved segmentation accuracy.
• To demonstrate the efficacy of a trainable pre-processing step using FCNs within the pipeline.

Main Methods:

• A novel pipeline combining Fully Convolutional Networks (FCNs) for pre-processing and Fully Convolutional Residual Networks (FC-ResNets) for segmentation.
• Utilizing FCNs to normalize medical input data before segmentation.
• Iterative refinement of segmentation predictions using the FC-ResNet component.

Main Results:

• Achieved state-of-the-art performance on Electron Microscopy benchmark datasets compared to 2D methods.
• Improved segmentation accuracy for CT liver lesions compared to standard FCN methods.
• Demonstrated competitive results on 3D MRI prostate segmentation, even against 3D methods, using the 2D pipeline.

Conclusions:

• The proposed FCN-FC-ResNet pipeline offers a powerful and versatile solution for medical image segmentation.
• Trainable pre-processing with FCNs enhances the performance of FC-ResNets.
• The pipeline shows strong potential for accurate segmentation across various medical imaging modalities and anatomical regions.