Segmentation of liver and liver lesions using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate organ and lesion segmentation in medical imaging is crucial for applications like nuclear medicine dosimetry and image analysis.
• Deep learning approaches for segmenting the liver and liver lesions on 3D MRI data remain an area requiring further investigation.

Purpose of the Study:

• To develop and evaluate a deep learning model for segmenting the liver and liver lesions using multi-contrast MRI data.
• To assess the performance of the model for potential applications in liver dosimetry and tumor detection.

Main Methods:

• Collected T1w and T2w MRI images from 128 patients, generating ground truth labels for the liver and lesions.
• Utilized a deep learning model based on the Isensee 2017 network, processing T1w and T2w images as a two-channel input.
• Evaluated model performance on 18 hold-out test datasets using the Dice coefficient.

Main Results:

• The deep learning model achieved an average Dice coefficient of 88% for liver segmentation across 18 test cases.
• Liver lesion segmentation yielded a lower average Dice coefficient of 53%.
• The model successfully segmented all 18 test cases, demonstrating high accuracy for liver delineation.

Conclusions:

• The developed deep learning model shows high accuracy for liver segmentation, suitable for applications such as liver dosimetry and attenuation correction in PET/MRI.
• The suboptimal performance in liver lesion segmentation indicates that the current method is not practical for tumor detection on clinical data.