■肝癌是最常见的癌症之一,基于能量的肿瘤消融是一种被广泛接受的治疗方法。肝肿瘤和消融区的自动和鲁棒分割将有助于治疗成功的评估。这项研究的目的是开发和评估一种基于自动深度学习的方法,用于(1)在动脉和门静脉阶段分割肝脏和肝脏肿瘤,用于治疗前CT,(2)在动脉和门静脉阶段分割肝脏和消融区,以进行消融治疗。
■252张CT图像纳入了在大型大学医院接受肝肿瘤消融术的63例患者;每位患者都有治疗前和治疗后的多期CT图像。肝脏的三维体素手动分割,肿瘤和消融区域由放射科医师提供参考标准。用于肝脏和病变分割的深度学习模型最初在公共肝脏肿瘤分割挑战(LiTS)数据集上训练以获得基础模型。然后,应用迁移学习来适应临床训练集上的基础模型,获取动脉和门静脉期图像的肿瘤和消融分割模型。对于建模,2D残余注意力Unet(RA-Unet)用于肝脏分割,基于多尺度贴片的3DRA-Unet用于肿瘤和消融分割。
■在独立测试装置上,所提出的方法在动脉和门静脉阶段的肝脏分割中实现了0.96和0.95的骰子相似系数(DSC),分别。对于肝脏肿瘤,动脉期模型检测灵敏度达71%,DSC为0.64,门静脉期灵敏度为82%,DSC为0.73。对于>0.5cm3的肝脏肿瘤,性能提高到79%的灵敏度,动脉期的DSC0.65,灵敏度86%,门静脉期DSC0.72。对于消融区,该模型对动脉期的检测灵敏度达到90%,DSC为0.83,对门静脉期敏感性为90%,0.89的DSC。
■所提出的深度学习方法可以在多相(动脉和门静脉)和多时间点(治疗前后)CT上提供肝肿瘤和消融区的自动分割,从而能够定量评估治疗成功。
OBJECTIVE: Liver cancer is one of the most commonly diagnosed cancer, and energy-based tumor ablation is a widely accepted treatment. Automatic and robust segmentation of liver tumors and ablation zones would facilitate the evaluation of treatment success. The purpose of this study was to develop and evaluate an automatic deep learning based method for (1) segmentation of liver and liver tumors in both arterial and portal venous phase for pre-treatment CT, and (2) segmentation of liver and ablation zones in both arterial and portal venous phase for after ablation treatment.
METHODS: 252 CT images from 63 patients undergoing liver tumor ablation at a large University Hospital were retrospectively included; each patient had pre-treatment and post-treatment multi-phase CT images. 3D voxel-wise manual segmentation of the liver, tumors and ablation region by the radiologist provided reference standard. Deep learning models for liver and lesion segmentation were initially trained on the public Liver Tumor Segmentation Challenge (LiTS) dataset to obtain base models. Then, transfer learning was applied to adapt the base models on the clinical training-set, to obtain tumor and ablation segmentation models both for arterial and portal venous phase images. For modeling, 2D residual-attention Unet (RA-Unet) was employed for liver segmentation and a multi-scale patch-based 3D RA-Unet for tumor and ablation segmentation.
RESULTS: On the independent test-set, the proposed method achieved a dice similarity coefficient (DSC) of 0.96 and 0.95 for liver segmentation on arterial and portal venous phase, respectively. For liver tumors, the model on arterial phase achieved detection sensitivity of 71%, DSC of 0.64, and on portal venous phase sensitivity of 82%, DSC of 0.73. For liver tumors >0.5cm3 performance improved to sensitivity 79%, DSC 0.65 on arterial phase and, sensitivity 86%, DSC 0.72 on portal venous phase. For ablation zone, the model on arterial phase achieved detection sensitivity of 90%, DSC of 0.83, and on portal venous phase sensitivity of 90%, DSC of 0.89.
CONCLUSIONS: The proposed deep learning approach can provide automated segmentation of liver tumors and ablation zones on multi-phase (arterial and portal venous) and multi-time-point (before and after treatment) CT enabling quantitative evaluation of treatment success.