An Optimized U-Net for Unbalanced Multi-Organ Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical practice is increasingly automating repetitive procedures to accelerate diagnosis.
• Semantic segmentation is crucial for identifying regions of interest in medical images by classifying pixels.
• Automated multi-organ segmentation faces challenges due to organ variability and computational demands.

Purpose of the Study:

• To develop an efficient, automated tool for semantic segmentation of abdominal organs.
• To address challenges in segmenting variable-sized organs and unbalanced data distributions.

Main Methods:

• Implementation of BIONET, a U-Net-based Fully Convolutional Network.
• Utilizing a U-Net architecture tailored for efficient semantic segmentation of abdominal organs.
• Addressing data imbalance issues inherent in physiological organ variations.

Main Results:

• BIONET achieved a Weighted Global Dice Score of 93.74 ± 1.1% for abdominal organ segmentation.
• Demonstrated high accuracy across variable organ dimensions with low variance.
• Achieved a fast inference performance of 138 frames per second.

Conclusions:

• BIONET provides a robust starting point for automated abdominal organ segmentation tools.
• The method effectively segments both small and large organs with high accuracy and low variability.
• This advancement supports faster and more standardized medical image analysis.