MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Multimodality Approach Using Deep Attention Convolutional Neural Networks for Localization of Intrahepatic Liver Cancer Recurrence Post-SBRT

L Wei*, I El Naqa, R Ten Haken, T Lawrence, University of Michigan, Ann Arbor, MI

Presentations

(Sunday, 7/12/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose:
To automatically acquire liver Couinaud segments and predict the location of intrahepatic recurrence segment-wise for primary Hepatocellular carcinoma (HCC) after SBRT treatment by combining multi-modality MR/CT images and 3D dose distributions.

Methods:
102 HCC patients with contrast-enhanced CT, T1 MRI, and 3D dose distributions were retrospectively analyzed. MR and 3D dose were co-registered to the planning CT. An unsupervised fast deformable registration neural network (VoxelMorph CNN) was trained to register a CT liver atlas to the planning CT. The 8 Couinaud liver segments for each patient were obtained automatically from the co-registered atlas CT. Stacked CECT, MR and 3D dose were fed into an attention deep CNN (ACNN) with focus on the primary tumor location and recurrent segment(s). The recurrence probability for each segment was estimated by averaging end-to-end the output score map. The ground truth of recurrence was multi-hot encoded for the liver segments with loss calculated by binary-cross-entropy-with-logits and optimized by Adam. The registration performance was measured by Dice coefficients and the Receiver Operating Characteristic curve was used for characterizing the performance of intrahepatic localization.

Results:
Training for both VoxelMorph and ACNN converged. The Dice coefficient for atlas to patient CT was 0.80 (std: 0.09). The registration and segmentation time for liver analysis was in the order of a few seconds once processed by trained VoxelMorph NN, much faster than conventional image registration algorithms. The ROC showed significant prediction performance of intrahepatic recurrence location over chance with an AUC of 0.709 (95% CI: 0.595-0.823).

Conclusion:
Unsupervised NN showed great potential for efficient and accurate liver deformable registration and Couinaud segmentation, especially for large datasets. The attention DCNN that embedded prior knowledge into its architecture showed promising performance for predicting intrahepatic recurrence and can provide future guidance for physicians for optimal management of complex intrahepatic treatment courses.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email