Versatile Convolutional Networks Applied to Computed Tomography and Magnetic Resonance Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnostics but faces challenges with diverse imaging modalities.
• Deep learning models are increasingly successful for automated medical image segmentation tasks.
• Tailored solutions are often required for different imaging types and anatomical regions.

Purpose of the Study:

• To demonstrate the versatility of a single deep learning architecture for medical image segmentation.
• To apply a unified model to distinct imaging modalities: computed tomography (CT) and magnetic resonance (MR) imaging.
• To validate the model's ability to segment various organs across different anatomical locations.

Main Methods:

• Development of a fully convolutional encoder-decoder deep learning architecture.
• Incorporation of high-resolution pathways for processing entire 3D volumes.
• Direct learning from data to identify and localize voxels belonging to regions of interest.

Main Results:

• The model achieved equivalent segmentation performance on both CT and MR imaging datasets.
• Successful segmentation of different organs in diverse anatomical regions was demonstrated.
• The architecture proved capable of processing whole 3D volumes efficiently.

Conclusions:

• A single deep learning architecture can effectively perform medical image segmentation across different modalities (CT and MR).
• The developed model exhibits versatility and robustness in segmenting various anatomical structures.
• This approach offers a promising unified solution for automated segmentation in medical imaging.