An effective automatic segmentation of abdominal adipose tissue using a convolution neural network.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Computational Anatomy

Background:

• Accurate segmentation of abdominal adipose tissue (AAT) is crucial for computer-aided diagnosis and prognosis using emission tomography images.
• Challenges include differentiating subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) due to subtle visual differences and complex distributions.
• Existing methods struggle with the multi-stage semantic segmentation required for VAT and SAT identification.

Purpose of the Study:

• To develop an automated convolutional neural network (A-CNN) for accurate segmentation of abdominal adipose tissue (AAT) from radiological images.
• To address the challenges in distinguishing between SAT and VAT in abdominal fat segmentation.
• To provide an efficient computational tool for AAT segmentation in clinical settings.

Main Methods:

• Development of an automated convolutional neural network (A-CNN) specifically designed for AAT segmentation.
• Implementation of a point-to-point learning process within the A-CNN.
• Utilization of a hybrid feature extraction technique for enhanced representation learning.

Main Results:

• The proposed A-CNN model demonstrated superior performance in AAT segmentation compared to existing deep learning methods on a CT dataset.
• The model achieved notable improvements in segmentation outcomes, particularly for the AAT region.
• Evaluation on a CT dataset confirmed the effectiveness of the A-CNN approach.

Conclusions:

• The developed A-CNN method is highly efficient and achieves remarkable performance, even on limited-scale, low-dose CT scans.
• The A-CNN presents a robust and fast computer-aided tool for clinical AAT segmentation.
• The study highlights the potential of A-CNN for improving diagnostic and prognostic capabilities in radiology.