Accurate monitoring of thermal ablation regions is an important guarantee for successful ablation treatment, which mainly depends on the subjective judgment of radiologists in current clinical practice. This work innovatively applied fully convolutional neural networks (FCNs) for detection and monitoring of thermal ablation regions in ultrasound (US) and comprehensively compared the performance of VGG16-FCN, U-Net, UNet++, Attention U-Net, MultiResUNet, and ResUNet, which have shown outstanding performance in medical image segmentation. The input of the models was US echo envelope data backscattered from the ablated regions. Excised porcine liver ablation dataset and clinical liver tumors ablation dataset were respectively used to evaluate the prediction ability of the models. With 1000 excised porcine liver ablation samples for training and 200 samples for testing, the UNet++ achieves both the highest Dice score (DSC) of 0.7824 ± 0.1098 and the best Hausdorff distance (HD) of 2.70 ± 1.38 mm. Additionally, considering potential clinical usage, we also tested the model generalizability by training on the excised dataset and testing on the clinical data, in which we obtained the performance with the highest DSC obtained by the ResUNet and the best HD by the UNet++. Our comparative study suggests that both UNet++ and ResUNet have relatively outstanding segmentation performance among all compared models, which are potential candidates for automatic segmentation of thermal ablation regions in US during clinical ablation treatment.Copyright © 2023 Elsevier B.V. All rights reserved.