Deep Learning Segmentation of Ascites on Abdominal CT Scans for Automatic Volume Quantification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Ascites is a common complication in patients with liver cirrhosis and ovarian cancer.
• Accurate quantification of ascites volume is crucial for treatment planning and monitoring.
• Current methods for ascites volume estimation can be time-consuming and subjective.

Purpose of the Study:

• To evaluate the performance of an automated deep learning (DL) method for detecting and quantifying ascites volume.
• To assess the DL model's accuracy in patients with liver cirrhosis and ovarian cancer using CT scans.

Main Methods:

• A retrospective study utilizing abdominal-pelvic CT scans from two institutions (NIH and UofW).
• A DL model was trained on The Cancer Genome Atlas Ovarian Cancer dataset and tested on internal (NIH-LC, NIH-OV) and external (UofW-LC) datasets.
• Performance was measured using Dice coefficient and median volume estimation error, compared against expert radiologist assessments.

Main Results:

• The DL model achieved high Dice scores (0.855±0.061 for NIH-LC, 0.826±0.153 for NIH-OV, 0.830±0.107 for UofW-LC).
• Median volume estimation errors were low (19.6% for NIH-LC, 5.3% for NIH-OV, 9.7% for UofW-LC).
• The model demonstrated strong agreement with expert assessments (r² values of 0.79, 0.98, and 0.97).

Conclusions:

• The automated deep learning method effectively segments and quantifies ascites volume.
• The DL model shows high accuracy and concordance with expert evaluations.
• This automated approach has the potential to improve clinical workflow for ascites management.