ESTRO 2020 Abstract book

S1049 ESTRO 2020

We split data into three sets: training, validation, and test. CNN was trained using each CT slice to increase the training data. The CNN architecture consisted of three convolutional layers with ReLU activation and kernel size 3, followed by a flattening layer and a final dense layer with 'softmax' activation for label classification. We used a mask for each OAR as input for the CNN.

the average radiotherapy planning time, calculated from localization tomography to the first fraction on the accelerator. The average value set at 21 days has recently decreased almost twice, but still 27.21% of patients in the analyzed period waited for the start of radiation therapy for more than 21 days. The next step was to determine the average time needed for the implementation of a single fraction, which resulted in the determination of the efficiency factor of the use of time on accelerators. The global average time of a single exposure in 2019 was 12.5 min, and the analysis of the time trend indicates slight fluctuations of this value over several months (picture 2). The global average exposure time is a variable function depending on parameters such as location, technique and type of accelerator.

Figure I shows the CNN.

By increasing our training set from one to 20 patients, we identified the minimum number needed to achieve a high CNN performance. The validation set was ten patients, whereas the test set changed from 82 (only one patient in the training set) to 63 (20 patients in the training set). For every CT slice, the CNN output was a label for each OAR contoured on that slice. For a whole CT, we sum all the slices' outputs and apply a Softmax function to normalize the final output. We chose to label the 3D-CT structure as the value that appeared more often in the outcome. We set a threshold of 70%. Below this value, we considered the CNN was not confident enough and labeled the structure as 'misclassification.' We analyze the accuracy of the output, excluding the misclassification structures, and the misclassification rate by increasing the training set. Results Accuracy of the CNN was already 100% with one patient. Figure 2 shows the misclassification rate and accuracy according to the number of patients in the training set. The misclassification rate decreases with the number of training patients. With a set of only 20 patients, the CNN achieves a misclassification rate of zero.

Conclusion Optimization of the radiotherapy process based on analytical tools that allow defining the patterns and trends that the collected data sets show is a continuous process. The analysis of time trends and the study of the interrelationships of a defined group of parameters gives the possibility of interference in real time and re- optimization of the process. PO-1788 Automated data labeling in radiotherapy using a deep neural network P.G. Franco 1 , E. Ambroa 2 , I. Valverde-Pascual 1 , J. Perez- Alija 1 1 Hospital de la Santa Creu i Sant Pau, Medical Physics, Barcelona, Spain ; 2 Consorci Sanitari de Terrassa, Medical Physics Unit- Radiation Oncology Department, Terrassa, Spain Purpose or Objective Deep learning models are mostly used in a supervised way, which require labeled data. Labeling data is a time- consuming process and a potential source of errors when done manually. The purpose of this study is to develop a convolutional neural network (CNN) to automatically identify and label the organs at risk (OAR) already contoured in a CT study. Material and Methods We collected data from 93 left breast patients. The contours used to train the CNN were: body (external contour), left lung, heart, breast PTV and lymph nodes PTV.

Conclusion