Fully convolutional network-based multi-output model for automatic segmentation of organs at risk in thorax.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of organs at risk (OARs) in thoracic computed tomography (CT) is crucial for radiation therapy planning.
• Manual delineation of OARs is time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop and evaluate a multi-output fully convolutional network (MOFCN) for the simultaneous and automatic segmentation of bilateral lung, heart, and spinal cord in thoracic CT slices.
• To assess the MOFCN's performance against established models and analyze its architectural components.

Main Methods:

• A MOFCN architecture was designed with a shared backbone and three distinct branches for segmenting lung, heart, and spinal cord.
• The model was trained and validated on 19,277 thoracic CT slices from 966 cancer patients, with expert radiation oncologists providing ground truth delineations.
• Performance was evaluated using the Dice similarity coefficient, and comparisons were made with other published models and variants to assess architectural impacts.

Main Results:

• The MOFCN achieved high segmentation accuracy, with Dice scores of 0.95 ± 0.02 for the lung, 0.91 ± 0.03 for the heart, and 0.87 ± 0.06 for the spinal cord.
• The MOFCN demonstrated comparable accuracy to other models while requiring significantly less time for segmentation.
• Analysis of model variants provided insights into the effectiveness of the distinct output design and the use of dilated convolutions.

Conclusions:

• The MOFCN effectively performs simultaneous and automatic segmentation of multiple OARs in thoracic CT scans.
• The model's efficiency in terms of parameters and computational requirements suggests a low hardware dependency.
• The MOFCN shows significant potential for widespread clinical application in radiation therapy planning.