Automatic abdominal multi-organ segmentation using deep convolutional neural network and time-implicit level sets.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Manual organ segmentation in CT scans is time-consuming and lacks reproducibility.
• Accurate segmentation is crucial for computer-aided diagnosis and surgical planning.

Purpose of the Study:

• To develop a fully automatic method for segmenting multiple organs from 3D abdominal CT images.
• To overcome the limitations of manual segmentation in terms of time and reproducibility.

Main Methods:

• Utilized deep fully convolutional neural networks (CNNs) for initial organ detection and segmentation.
• Employed a time-implicit multi-phase evolution method for refining segmentation accuracy.
• Incorporated image intensity models, probability priors, and disjoint region constraints into an energy functional, optimized via a novel level-set algorithm.

Main Results:

• Achieved high accuracy with average Dice overlap ratios of 96.0% (liver), 94.2% (spleen), and 95.4% (kidneys) on 140 CT scans.
• Demonstrated an average symmetric surface distance of less than 1.3 mm for all segmented organs.
• Completed segmentation of a CT volume in an average of 125 seconds, showing superior efficiency compared to state-of-the-art methods.

Conclusions:

• Developed and validated a fully automatic multi-organ segmentation method for abdominal CT images.
• The method shows significant potential for clinical use due to its high effectiveness, robustness, and efficiency.