CT based automatic clinical target volume delineation using a dense-fully connected convolution network for cervical
Zhongjian Ju1, Wen Guo2,3, Shanshan Gu1
1Department of Radiation Oncology, The First Medical Center, People's Liberation Army General Hospital, No. 28 Fuxing Road, Haidian District, Beijing, 100853, China.
BMC Cancer
|March 9, 2021
Summary
Accurate cervical cancer radiotherapy requires precise Clinical Target Volume (CTV) delineation. Dense V-Net effectively automates CTV pre-delineation on CT scans, improving accuracy and efficiency for patient treatment.
Area of Science:
- Medical Imaging
- Radiotherapy
- Artificial Intelligence
Background:
- Accurate delineation of the Clinical Target Volume (CTV) on patient CT images is crucial for effective radiotherapy.
- Current manual delineation methods are hindered by limited clinical samples and difficulties in automation, slowing research for cervical cancer patients.
- This study investigates the potential of a Dense-Fully Connected Convolution Network (Dense V-Net) for automated CTV pre-delineation in cervical cancer radiotherapy.
Purpose of the Study:
- To evaluate the efficacy of Dense V-Net in predicting the CTV pre-delineation for cervical cancer patients undergoing radiotherapy.
- To assess the performance of Dense V-Net in automating the complex task of CTV delineation using CT images.
Main Methods:
- A Dense V-Net model was employed for automatic CTV pre-delineation on computed tomography (CT) images of 133 cervical cancer patients (Stage IB-IIA, postoperative).
- The dataset was divided into a training set (113 patients) for model parameter adjustment and a test set (20 patients) for performance evaluation.
- Pre-sketching accuracy was assessed using 8 representative parameters, including sketching similarity, offset, and volume difference.
Main Results:
- The Dense V-Net achieved high accuracy in CTV pre-delineation, with key metrics such as Dice Similarity Coefficient (DSC) of 0.82 ± 0.03 and Hausdorff Distance (HD/cm) of 1.86 ± 0.48.
- Other evaluated parameters including DC/mm, MAD/mm, ∆V, SI, IncI, and JD also demonstrated favorable performance.
- The results indicated superior performance compared to single network approaches.
Conclusions:
- Dense V-Net demonstrates significant capability in accurately predicting CTV pre-delineation for cervical cancer patients.
- The model shows promise for clinical application in radiotherapy, potentially requiring only minor modifications for implementation.
