Focal liver lesion diagnosis with deep learning and multistage CT imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate diagnosis of liver lesions is critical for effective treatment and patient prognosis.
• Multiphase enhanced computed tomography (CT) is a key imaging modality for liver lesion evaluation.

Purpose of the Study:

• To develop and validate an automated system, the Liver Lesion Network (LiLNet), for the diagnosis of focal liver lesions using multiphase enhanced CT.
• To assess the diagnostic performance of LiLNet in classifying various types of liver lesions, including both benign and malignant tumors.

Main Methods:

• Development of the Liver Lesion Network (LiLNet) using a dataset of 4039 patients from six data centers.
• LiLNet was trained to identify and classify six types of focal liver lesions: hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma (ICC), metastatic tumors (MET), focal nodular hyperplasia (FNH), hemangioma (HEM), and cysts (CYST).
• Validation was performed across four external centers and clinical verification in two hospitals.

Main Results:

• LiLNet achieved high diagnostic accuracy: 94.7% ACC and 97.2% AUC for differentiating benign and malignant tumors.
• For malignant tumors (HCC, ICC, MET), ACC was 88.7% and AUC was 95.6%.
• For benign lesions (FNH, HEM, CYST), ACC was 88.6% and AUC was 95.9%.

Conclusions:

• The developed Liver Lesion Network (LiLNet) demonstrates significant potential as an automated tool for liver lesion diagnosis.
• LiLNet can aid clinicians in diagnosis, particularly in resource-limited settings with a shortage of expert radiologists.
• The system shows high accuracy in classifying both benign and malignant liver lesions, supporting clinical decision-making.