Quantitative Automated Segmentation of Lipiodol Deposits on Cone Beam CT Imaging acquired during Transarterial Chemoembolization for Liver Tumors: A Deep Learning Approach.

Researchers

Journal

Modalities

Models

Abstract

The purpose of this study was to show that a deep learning-based, automated model for Lipiodol segmentation on CBCT after cTACE performs closer to the “ground truth segmentation” than a conventional thresholding-based model.This post-hoc analysis included 36 patients with a diagnosis of HCC or other solid liver tumor who underwent cTACE with an intra-procedural CBCT. Semi-automatic segmentation of Lipiodol were obtained. Then, a convolutional U-net model was used to output a binary mask that predicts Lipiodol deposition. A threshold value of signal intensity on CBCT was used to obtain a Lipiodol mask for comparison. Dice similarity coefficient (DSC), Mean-squared error (MSE), and Center of Mass (CM), and fractional volume ratios for both masks were obtained by comparing them to the ground truth (radiologist segmented Lipiodol deposits) to obtain accuracy metrics for the two masks. These results were used to compare the model vs. the threshold technique.For all metrics, the U-net outperformed the threshold technique: DSC (0.65±0.17 vs. 0.45±0.22,p<0.001) and MSE (125.53±107.36 vs. 185.98±93.82,p=0.005). Difference between the CM predicted, and the actual CM was (15.31±14.63mm vs. 31.34±30.24mm,p<0.001), with lesser distance indicating higher accuracy. The fraction of volume present ([predicted Lipiodol volume]/[ground truth Lipiodol volume]) was 1.22±0.84vs.2.58±3.52,p=0.048 for our model’s prediction and threshold technique, respectively.This study showed that a deep learning framework could detect Lipiodol in CBCT imaging and was capable of outperforming the conventionally used thresholding technique over several metrics. Further optimization will allow for more accurate, quantitative predictions of Lipiodol depositions intra-procedurally.Copyright © 2021. Published by Elsevier Inc.