CT-based artificial intelligence system complementing deep learning model and radiologist for liver fibrosis staging.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Hepatology

Background:

• Liver fibrosis staging is critical for predicting patient morbidity and mortality.
• Noninvasive methods for liver fibrosis staging are in high demand.
• Current methods face challenges in accuracy and generalizability.

Purpose of the Study:

• To develop an automated deep learning (DL) model for noninvasive liver fibrosis staging.
• To create a deep learning-radiologist complementarity decision system (DRCDS) for improved diagnostic accuracy.
• To address model generalization and human-machine complementarity in multi-classification tasks.

Main Methods:

• Developed an automated DL-based segmentation and classification model (Model-C).
• Employed test-time adaptation to mitigate data distribution shifts.
• Established a DRCDS using a decision model for AI-radiologist collaboration.

Main Results:

• Model-C demonstrated high performance (AUCs of 0.89-0.92), outperforming liver-only or spleen-only models.
• Test-time adaptation improved Model-C's Obuchowski index on external datasets.
• DRCDS slightly outperformed Model-C and senior radiologists, with high adoption rates of Model-C's diagnosis (73.7%-92.0%).

Conclusions:

• The DRCDS offers a highly accurate approach for diagnosing liver fibrosis.
• The study provides effective solutions for model generalization and human-machine complementarity in medical AI.
• This work advances noninvasive liver fibrosis staging and AI-assisted clinical decision-making.