Artificial intelligence-based detection and assessment of ascites on CT scans

Clinically important ascites are a result of multifactorial pathogenesis. Planning therapy merely depends on precisely detecting and quantitatively classifying ascites to minimize potential adverse effects. However, manually segmenting and quantifying ascites is time-consuming asascites typically appear on multiple CT scans. In this study, an AI-based approach (SRFLab) is developed to quantify ascites from CT scans automatically. First, abdomen sections are automatically acquired from the retrospectively screened CT volume using multitask classification (AcquNet). The proposed CNN is retrieved under a task-specific objective using transfer learning. Alternatively, ascites are learned from a supervision representation fusion CNN (QuanNet) to evaluate fluid formation. Experimental results demonstrate that the proposed schema leads to good performance compared to other existing methods. AcquNet achieved a mean accuracy of 97.80% +/- and a 1.97% standard deviation, while the accuracy of QuanNet achieved a mean accuracy of 97.21% +/- and a 2.61% standard deviation. Overall, the results of this study demonstrate the effec-tiveness of the proposed model and the advancement of the volumetric assessment of ascites on CT volume images. The proposed model is more efficient at detecting and quantifying ascites in patients than clinical experts. Thus, the proposed model can support the rapid grading of ascites on CT volume images and aid radiologists in clinical practice.