Site Editor



ASTRO 2022: Deep Learning–Based Framework in Liver Tumor Localization

By: Vanessa A. Carter, BS
Posted: Thursday, October 27, 2022

Hua-Chieh Shao, PhD, of The University of Texas Southwestern Medical Center, Dallas, and colleagues aimed to track real-time liver tumor localization using a deep learning–based framework that explored correlations between patients’ internal anatomic motion and body surface maps. Dr. Shao presented these findings during the 2022 American Society for Radiation Oncology (ASTRO) Annual Meeting (Abstract 154), concluding that this technology “allows accurate 3D liver tumor tracking in real-time via nonionizing, marker-less, and high frame–rate optical surface imaging.”

Real-time liver tumor localization was established by combining intraliver motion propagation via deep learning–based biomechanical modeling from liver boundary motion and liver boundary motion estimation via deep features learned from optical body surface imaging. A patient-specific, fully connected convolutional neural network (SurfCNN) predicted the motion of nodes by using real-time optical body surface imaging, and a U-Net-style model (UNet-Bio) inspired by biomechanical modeling used the liver boundary motion to infer intraliver tumor motion.

This framework was assessed using data from eight patients with liver cancer. The data set of each patient was altered by simulating real-time motion patterns to test the SurfCNN model. A total of 1,728 different motion states were observed by analyzing each patient’s four-dimensional CT scans. UNet-Bio may be trained as a population-based model that includes data from intraliver deformation vector fields solved by finite element analysis–based biomechanical modeling.

Using a total of 576 unseen scenarios for each patient case, the cascaded SurfCNN and UNet-Bio framework was able to localize liver tumors to an average of 0.791 ± 0.141 in dice similarity coefficient, 3.6 ± 1.9 mm in Hausdorff distance, and a 2.3 ± 1.7 mm center-of-mass error. For reference, the prior image without deformable registration yielded an average of 0.534 ± 0.277 in dice similarity coefficient, 8.3 ± 6.1 mm in Hausdorff distance, and a 6.8 ± 5.8 center-of-mass error. Of note, the whole-model inference time was less than 100 msec.

Disclosure: Dr. Shao reported no financial conflicts of interest.

By continuing to browse this site you permit us and our partners to place identification cookies on your browser and agree to our use of cookies to identify you for marketing. Read our Privacy Policy to learn more.