ESTRO meets Asia 2024 - Abstract Book

S319

Physics – Machine learning models and clinical applications

ESTRO meets Asia 2024

174

Proffered Paper

development of deep learning based delineation method of clinical target volume for lung tumor

Shota Ishihara 1,2 , Akihiro Takemura 2 , Naoki Hayashi 3 , Katsuhiko Otobe 1 , Fumihiko Niwa 1

1 Department of Medical Technology, Ogaki Municipal Hospital, Gifu, Japan. 2 Faculty of Health Sciences, Institute of Medical, Pharmaceutical and Health Sciences, Kanazawa University, Ishikawa, Japan. 3 Division of Medical Physics, School of Medical Sciences, Fujita Health University, Aichi, Japan

Purpose/Objective:

Radiotherapy for lung tumors, such as stereotactic body radiation therapy (SBRT), requires accurate delineation of internal target volume (ITV) and clinical target volume (CTV) because the accuracy of delineation would affect treatment outcomes. ITV is usually calculated from CTVs in maximum intensity projection images of four dimensional computed tomography (4DCT-MIP) images, which depicts a range of tumor respiratory motion and planning computed tomography (pCT) images are usually used for treatment planning to delineate CTV and the other target volumes. However, delineation of the CTVs for each image is laborious, time-consuming, and has an issue of inter-observer variability. It was reported that auto-CTV delineation with deep learning was efficient and could provide less variation of CTV contour. An auto-CTV delineation succeeded to assist delineation of CTV 1) . However, there is no auto-CTV delineation methods appliable to both pCT and 4DCT-MIP which are acquired for SBRT. The objectives of this paper was to develop a deep learning based automatic delineation method for CTV in pCT images and 4DCT-MIP images, and to validate its delineation accuracy. The deep learning model we used for this study was based on U-Net and included residual blocks. The public datasets “The Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) 2) ,” provided by The Cancer Imaging Archive was used for the training for this model. From the LIDC-IDRI dataset, 1350 images from 90 cases were selected and divided into training (n = 1000), validation (n = 250), and test (n = 100) datasets. The tumor region was cropped to 96 pixels ×96 pixels from the images. Three types of data augmentation, rotation, flipping, and translation, were applied. Training was performed with a batch size of 100, and the number of epochs was set to 100. Early termination was also implemented to prevent overfitting. Adam optimizer was used with a learning rate of 10 -4 , and the Dice loss was used as the loss function. To validate the performance of our auto-CTV delineation method for lung tumor, other 110 images of the pCT and the 4DCT-MIP which were acquired from 10 cases were used. The performance of the auto-CTV delineation method for the pCT and the 4DCT-MIP were measured by using three indices: dice similarity coefficient (DSC), precision, and recall. The DSC for the pCT images was statistically compared with that for the 4DCT-MIP images by t-test. Material/Methods:

Results:

Training was interrupted in 74 epochs to avoid the risk of overfitting. At that epoch the dice loss was 0.16, and the DSC, precision, and recall were 0.81, 0.82, and 0.81 for the validation dataset, respectively, and were 0.87, 0.90, and 0.84 for the test dataset, respectively. Those for the pCT dataset were 0.62, 0.66, and 0.65, respectively. However, those for the 4DCT-MIP were 0.53, 0.59 and 0.55, respectively and were lower than those for the pCT. In both results for the pCT and the 4DCT-MIP, the areas which the auto-CTV delineation method output tended to be smaller than the label data. DSC for the pCT was significant different from that for the 4DCT-MIP (p<0.05) 3) .