RT Journal Article SR Electronic T1 Performance assessment of a pre-trained CNN model for auto-delineation of primary tutors on CT scans in non-small cell lung carcinoma (NSCLC) patients JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 3201 OP 3201 VO 63 IS supplement 2 A1 Jha, Ashish A1 Kulkarni, Chaitanya A1 Mithun, Sneha A1 Sherkhane, Umesh kumar A1 jaiswar, Vinay A1 Mehta, Grace Monica S. A1 Panchal, Shreyash A1 Nath, Biprojit A1 Nautiyal, Amit A1 M.S., Dinesh A1 PURANDARE, NILENDU A1 RANGARAJAN, VENKATESH A1 Wee, Leonard A1 Dekker, Andre YR 2022 UL http://jnm.snmjournals.org/content/63/supplement_2/3201.abstract AB 3201 Introduction: Computed Tomography (CT) is used in the diagnosis, planning, and follow-up of disease in cancer treatment. In the last few years, CT-based radiomic features are being researched in precision oncology to predict the various clinical endpoints. Delineation of tumor in the medical images is an important step in cancer treatment by radiotherapy and also in the extraction of radiomic features from tumor. Usually, tumor delineation is performed manually by an imaging expert or a treating radiation oncologist which is operator-dependent. A slight change in tumor delineation may impact radiomic extraction and cause radiomic feature instability. This study aims to develop a deep learning model to leverage this advantage to automate the delineation of tumor in lung cancerMethods: The study is approved by the institutional ethics committee (IEC) retrospective study with a waiver of informed consent. A total of 197 non-small cell lung carcinoma (nsclc) patients’ pre-treatment CT scans were used for this study. Tumor delineation was performed by two 15 years experienced imaging medical physicists and reviewed by 30 years experienced nuclear medicine physician. A pre-trained convoluted neural network (CNN) model on 421 nsclc patients’ pre-treatment CT scan (Lung-1 dataset, Maastro clinic, Netherlands) was used for this study. The study was split into two parts a) Study1: direct validation of the model on our data set and b) Study2: retraining and validation of model on our data set. Study1: The pretrained CNN model was validated on 100 patients’ data sets. Study2: The data was split into training and validation sets (a 97/104). The pretrained CNN model was retrained on the training set and validated of the validation set. Model performance was accessed based on dice score, precision, and recall calculated on the validation set of both the studies independently. Results: In study1 the median dice score, precision, and recall on the validation set were found to be .57, 0.61, 0.54 respectively. In Study2 the median dice score, precision, and recall on the validation set were 0.774, 0.854, and 704 respectively. The detail of the validation score of both the studies is shown in table 1. The comparative image of manual and automatic delineation is shown in figure 1. Table1: shows the model performance score in the validation set in both the studies Study1 and 2.Conclusions: However, a pretrained CNN model did not show the optimum performance in the delineation of primary tumor of nsclc patients in the external validation set. But with this study, we were able to demonstrate that the retraining of the pre-trained CNN model improves the performance of the model in the delineation of the primary tumor in nsclc patients.