ESTRO meets Asia 2024 - Abstract Book

S320

Physics – Machine learning models and clinical applications

ESTRO meets Asia 2024

Conclusion:

The auto-CTV delineation method well performed for the test dataset from LIDC-IDRI, but the performance of the method was slightly decreased for the pCT data and the 4DCT-MIP data. The DSC for the pCT was better than the DSC for the 4DCT-MIP. The LIDC-IDRI dataset contains various CT image data and the image quality is similar to that of the pCT images but is not similar to that of the 4DCT-MIP images. Therefore, the model trained with the LIDC-IDRI dataset provided lower DSC for the 4DCT-MIP.

Keywords: SBRT, deep learning, 4DCT

References:

1)Deep learning improved clinical target volume contouring quality and efficiency for postoperative radiation therapy in non-small cell lung cancer; Nan Bi, et al.; Frontiers in oncology; 2019; 9; 1192 2) The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A completed reference database of lung nodules on CT scans.; Armato SG 3rd, et al.; Medical Physics; 2011; 38; 915-931 3)Automated approach for segmenting gross tumor volumes for lung cancer stereotactic body radiation therapy using CT-based dense V-networks; Yunhao Cui, et al.; Journal of radiation research; 2021; 62; 346-355

314

Digital Poster

Development of an In-House Deep Learning Model for Prostate Segmentation in Radiotherapy Planning

Diyana Afrina Hizam 1 , Tan Li Kuo 2 , Marniza Saad 1 , Ung Ngie Min 1

1 Clinical Oncology Unit, University of Malaya, Kuala Lumpur, Malaysia. 2 Department of Biomedical Imaging, University of Malaya, Kuala Lumpur, Malaysia

Purpose/Objective:

This study aimed to develop and evaluate an in-house deep learning model for automatic segmentation of the prostate and surrounding structures relevant to radiotherapy planning, including seminal vesicles (SV) and the penile bulb (PB). These structures are often excluded in current deep learning-based segmentation studies and are absent in the publicly available dataset.

Material/Methods:

A publicly available dataset (1) containing 131 prostate cancer cases, comprising CT images, was utilized for initial training. It’s worth noting that this is the only publicly available CT image dataset for prostate cancer radiotherapy. This dataset included segmentations for the bladder, femoral heads, rectum, and prostate. The model, a 2D U Net architecture, was subsequently fine-tuned on a local dataset of 30 cases encompassing all structures from the public dataset, along with SV and PB. Both quantitative and qualitative evaluations were performed to assess the model's segmentation accuracy for all structures (2).

Results:

The developed model demonstrated good segmentation performance for all structures, including the SV and PB, despite their smaller volume compared to other structures. These results are supported by both quantitative