Deep learning-enhanced 3D echo decorrelation imaging for monitoring radiofrequency ablation in ex vivo human liver

Three-dimensional (3D) ultrasound echo decorrelation imaging can successfully monitor treatment of liver tumors by radiofrequency ablation (RFA), but has limitations in mapping ablation zones and tissue temperature. Here, supervised deep learning was investigated to improve prediction of temperature and ablation from 3D echo decorrelation images. RFA was performed on N = 71 specimens of ex vivo normal, steatotic, and cirrhotic human liver tissue. During ablation, echo volume pairs were acquired with a 4.5 MHz matrix array, 3D echo decorrelation images were computed, and temperatures were measured by 4 thermocouples integrated into the RFA probe. For temperature prediction, a fully connected neural network was trained on local echo decorrelation to minimize mean-squared-error vs. measured temperatures. Predictions were significantly correlated with measurements (r = 0.842, p < 10⁻¹⁶), with overall root-mean-square error 17.28 °C. After RFA, scanned tissue sections were manually segmented to serve as ground truth for ablation zone prediction. 3D U-Net convolutional neural networks processing cumulative echo decorrelation volumes predicted ablation zones significantly better than thresholding of decorrelation maps, based on areas under receiver-operating-characteristic curves (0.988 vs. 0.946) and precision-recall curves (0.616 vs. 0.430), average Dice coefficients between predicted and measured ablation zones (0.622 vs. 0.396), and normalized root-mean-squared error of volume prediction (27.1% vs. 35.3%).