Body Location Embedded 3D U-Net (BLE-U-Net) for Ovarian Cancer Ascites Segmentation on CT scans

Area of Science:

• Medical Imaging
• Oncology
• Artificial Intelligence

Background:

• Ascites is a critical indicator of advanced ovarian cancer, the deadliest gynecologic malignancy.
• Accurate ascites measurement is vital for monitoring disease progression and guiding treatment.
• Segmentation is difficult due to similar-looking fluids near ascites in CT scans.

Purpose of the Study:

• To develop an improved method for segmenting ascites in ovarian cancer patients.
• To enhance the accuracy of ascites measurement using deep learning techniques.
• To investigate the impact of anatomical location information on segmentation performance.

Main Methods:

• A novel 3D U-Net segmentation model, BLE-U-Net, was proposed.
• The model integrates body part regression to predict anatomical location along the z-axis.
• Regression scores were discretized and embedded into the 3D U-Net architecture.

Main Results:

• BLE-U-Net achieved a Dice coefficient of 65 ±06, significantly outperforming a conventional 3D U-Net (38 ±10).
• Segmented ascites volumes closely matched ground truth (0.57±0.85 L vs 0.58±0.84 L).
• The improvement was attributed to the integration of location information (p <0.05).

Conclusions:

• Integrating anatomical location information significantly improves ascites segmentation accuracy.
• BLE-U-Net shows potential for enhancing ovarian cancer diagnosis and treatment monitoring.
• This AI-driven approach offers a promising tool for managing ascites in gynecologic malignancies.