Estimating 3-dimensional liver motion using deep learning and 2-dimensional ultrasound images

Area of Science:

• Medical imaging
• Surgical technology
• Computational anatomy

Background:

• Current tumor tracking systems for radiofrequency ablation (RFA) are limited to 2D ultrasound (US) images, failing to account for 3D organ motion.
• This limitation can lead to loss of the ablation area and reduced tumor visibility during RFA procedures.
• Accurate estimation of 3D organ movement is crucial for effective tumor tracking in RFA.

Purpose of the Study:

• To develop a system for tracking tumor position during RFA by estimating 3D liver movement.
• To improve the accuracy of 3D liver movement estimation by enhancing liver segmentation in US images.
• To enable precise estimation of the relative 6-axial movement between the liver and the US probe.

Main Methods:

• Utilized a convolutional neural network (CNN) for 3D displacement estimation from 2D US images.
• Integrated liver segmentation maps as input to a regression network to enhance movement estimation accuracy.
• Developed a bi-directional convolutional LSTM U-Net with densely connected convolutions (BCDU-Net) for improved liver segmentation.

Main Results:

• The BCDU-Net significantly improved liver segmentation accuracy.
• Enhanced segmentation accuracy led to a corresponding improvement in 3D liver movement estimation.
• Achieved a mean absolute error of 0.0645 mm/frame for out-of-plane movement estimation.

Conclusions:

• The proposed BCDU-Net and CNN method effectively identifies liver movement for RFA tumor tracking.
• Precise liver segmentation using BCDU-Net is key to enhancing the performance of liver movement estimation.
• This approach offers a promising solution for improving tumor tracking accuracy in RFA treatments.