Deep learning for abdominal adipose tissue segmentation with few labelled samples.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate segmentation of abdominal adipose tissue, including subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT), is crucial for biomedical diagnoses and prognoses.
• Traditional clinical methods for SAT and VAT segmentation are labor-intensive, costly, time-consuming, and prone to segmentation errors.

Purpose of the Study:

• To introduce and develop EFNet, an effective convolutional neural network (ConvNet) for fully automated abdominal adipose tissue segmentation from CT scans.
• To address the challenges of multistage semantic segmentation and high intensity similarity between VAT and SAT using a novel deep learning approach.

Main Methods:

• EFNet employs a three-pathway architecture: max unpooling for computational efficiency, concatenation for shape recovery, and anatomy pyramid pooling for fine-grained feature extraction.
• The model encodes usable anatomical information, allowing control over the density of fine-grained features.
• An end-to-end learning process with a mixed feature fusion layer jointly learns representation features.

Main Results:

• EFNet demonstrated superior performance compared to existing deep learning networks in abdominal adipose tissue segmentation.
• The model achieved significant improvements in segmentation accuracy and efficiency on diverse datasets.
• Evaluations confirmed EFNet's tremendous performance in segmenting abdominal adipose tissue.

Conclusions:

• EFNet provides extremely fast and remarkably accurate fully automated segmentation of VAT and SAT from CT scans.
• The proposed method shows strong potential for enhancing automated detection and segmentation of abdominal adipose tissue in clinical practice.