Automatic Delineation of the Hippocampus Based On Deep Learning Using Multimodal Imaging for Cerebral Radiotherapy
R Zhang1*, Y Liu2, C Fu3, J Zhang4, D Li5, C Chen6, Z Dai7, L Jia8, W Zhang9, (1) The First Hospital of Tsinghua University, Beijing, CN (2) Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, CN (3) The First Hospital of Tsinghua University, Beijing, CN (4) The First Hospital of Tsinghua University, Beijing, CN (5) The First Hospital of Tsinghua University, Beijing, CN (6) The First Hospital of Tsinghua University, Beijing, CN (7) United Imaging Research Institute of Intelligent Imaging, Beijing, CN (8) Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, CN (9) Shanghai United Imaging Healthcare CO., Ltd, Shanghai, CN
Presentations
MO-E115-BOP-F1-1 (Monday, 7/11/2022) 1:15 PM - 1:45 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 1
Purpose: In order to be able to minimize dose deposition within the hippocampus in whole brain irradiation, clinicians must be able to confidently delineate the hippocampus structure. The technique of efficient and accurate delineation of the hippocampus to provide the precise protection for the structure in cerebral radiotherapy is extremely important.
Methods: In this work, CT and MRI of 40 patients with brain metastases treated in our department from January 2020 to December 2020 were collected retrospectively. An automatic delineation model (3D BUC-Net) based on deep learning was designed by combining 3D bottler layer module and the cascade architecture to improve the accuracy and efficient of the hippocampus delineation and respectively trained and evaluated on two datasets, CT and CT-MRI. Our model was evaluated by comparing Dice Similarity Coefficient (DSC), 95% Harsdorf Distance (95HD) and absolute error rate (AER) of the volume of two hippocampus contours delineated automatically by the models and delineated manually by clinicians with 3D U-Net and 3D U-Net Cascade. The delineation efficiency was evaluated by the time-consuming that a patch of 3D imaging is delineated automatically.
Results: The delineation accuracy of the hippocampus for three models on the dataset CT-MRI is significantly superior to the dataset CT; BUC-Net achieves the best delineation accuracy for left hippocampus (DSC: 0.900, 95HD: 0.792 mm, AER: 2.0%) and right hippocampus (DSC: 0.882, 95HD: 0.823 mm, AER: 3.1%) on CT-MRI dataset. Moreover, the delineation efficiency of BUC-Net outperforms that of other two models.
Conclusion: BUC-Net can achieve more efficient and accurate delineation for the hippocampus on multimodal imaging combined CT with MRI, which provides a lot of convenience for the protection of the hippocampus during cerebral radiotherapy.
Keywords
Not Applicable / None Entered.
Taxonomy
IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking
Contact Email