Monitoring radiofrequency ablation in ex vivo human liver using 3D echo decorrelation imaging augmented by deep learning

Monitoring of radiofrequency ablation (RFA) is desirable to improve safety and efficacy of liver tumor treatment. Three-dimensional ultrasound echo decorrelation imaging can successfully predict local ablation effects but has had limited success in mapping ablation zone margins and local tissue temperature. Here, a supervised deep learning approach is investigated to improve prediction of ablation zones and tissue temperature from 3D echo decorrelation images. RFA was performed on ex vivo human liver tissue, including normal, fibrotic, and cirrhotic liver (N > 30). During ablation, pairs of echo volumes were acquired with a 4.5 MHz transesophageal matrix array, 3D echo decorrelation images were computed for each volume pair, and temperatures measured by fourthermocouples integrated into the RFA probe were recorded. Tissue was then frozen, sectioned, scanned, and ablation zones were manually segmented. For prediction of ablation zones, B-mode and echo decorrelation images were input to a U-net convolutional neural network to segment ablation margins, with histology serving as ground truth for training and cross-validation. For prediction of temperature, echo decorrelation values at the thermocouple locations were used as input to train a dense network, minimizing mean-squared-error versus measured temperatures. The results indicate promise for improved mapping of tissue ablation and temperature.