3D Deep Learning for Anatomical Structure Segmentation in Multiple Imaging Modalities.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate segmentation of anatomical structures in radiological scans (MRI, CT) is crucial for biomarkers and computer-aided diagnosis (CADx).
• Existing automated segmentation methods face challenges due to anatomical variations, imaging artifacts, and motion artifacts from breathing.
• Robust segmentation is essential for reliable interpretation of multi-protocol medical images.

Purpose of the Study:

• To present a novel deep learning approach for automatic organ and muscle segmentation in multi-modal medical images.
• To address challenges in automated segmentation, including anatomical variability and motion artifacts.
• To improve the accuracy and detail of segmentation for enhanced diagnostic support.

Main Methods:

• A two-part deep learning process utilizing a 3D encoder-decoder (Rb-UNet) for localization and a 3D Tiramisu network for boundary-preserving segmentation.
• The Rb-UNet predicts a 3D bounding box for the target structure.
• The Tiramisu model then performs detailed segmentation within the predicted bounding box.

Main Results:

• The proposed method was evaluated on six diverse datasets (MRI, DCE-MRI, CT) including pancreas, liver, kidneys, and psoas muscle.
• Achieved mean Dice Similarity Coefficient (DSC) scores comparable to or surpassing state-of-the-art methods.
• Qualitative assessment by radiologists confirmed the preservation of detailed organ and muscle boundaries.

Conclusions:

• The novel deep learning approach effectively performs accurate and detailed organ and muscle segmentation in multi-modal medical imaging.
• This method offers a robust solution to challenges in automated segmentation, enhancing its utility in clinical settings.
• The validated accuracy and boundary preservation support its integration into computer-aided diagnosis systems.