Deep learning for noninvasive liver fibrosis classification: A systematic review

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Liver biopsy, the standard for fibrosis staging, carries significant risks.
• Noninvasive liver fibrosis assessment is a rapidly advancing field.
• Deep learning (DL) is transforming medical image analysis, offering potential for improved fibrosis evaluation.

Purpose of the Study:

• To systematically review the application of deep learning in noninvasive liver fibrosis imaging.
• To assess the accuracy and limitations of DL in classifying liver fibrosis.

Main Methods:

• A systematic literature search was conducted across Embase, MEDLINE, Web of Science, and IEEE Xplore.
• Keywords included "liver/hepatic," "fibrosis/cirrhosis," and "neural/DL networks."
• Risk of bias and applicability were assessed using the QUADAS-2 tool.

Main Results:

• Sixteen studies utilizing ultrasound, CT, and MRI were analyzed, involving 15,853 patients.
• Most studies reported high accuracy (>85%) compared to histopathology.
• However, 87.5% of studies exhibited a high risk of bias and applicability concerns, with most being retrospective.

Conclusions:

• Deep learning holds potential for classifying liver fibrosis noninvasively.
• The field is currently limited by a scarcity of prospective studies and data accessibility.
• Standardized reporting and database sharing are crucial for optimizing DL in large-scale liver fibrosis evaluation.