ESTRO 2021 Abstract Book

S1415

ESTRO 2021

Organs at risk (OAR’s) mistakenly depicted at the time of the treatment plan cause a decrease in the accuracy of the treatment plan. This study verifies whether the OAR’s circular contour errors could be automatically detected using a convolutional neural network (CNN). Materials and Methods Taking OAR’s (heart and lungs) depicted using 135 chest non-small cell lung cancer (NSCLC) CT images from The Cancer Imaging Archive, we drew circles with a diameter of 8 and 12 mm in a position 10 mm from each OAR’s limbic on the CT slice, at random, as circular contour errors. Using a DICOM RT Structure Set, we generated images of OAR’s masks, including OAR’s depicted normally and circular contour errors. The mask images were created by resizing 512 × 512 px chest CT images to 128 × 128 px. CNNs are comprised of 1 input layer, 8 convolutional layers (Conv1 ～ 8), and 4 fully connected layers (FC1 to FC3+output layer), and there is a shortcut connection between Conv3 and 5, and Conv6 to 8 (Figure 1). With one maximum pooling layer added to the Conv2, 5, 8 output, batch normalization was performed on each. The filter sizes for the convolutional layer and maximum pooling layer were 3×3 and 2×2, respectively. Binary classification, with and without errors, was then performed with the fully-connected layer output layer. For the activation functions, Flexible rectified linear unit (FReLU) was used in the convolutional layers, and Mish and Softmax were used in the fully-connected layers (FC1 ～ 3). Optimizer used AdaBelief. The loss function used cross entropy. For the number of epochs, there was processing to automatically stop when an increase in the accuracy of the loss function was not observed between 5 epochs. Flip and GridMask were used for data augmentation. The accuracy evaluation divided the mask images into training data and validation data and performed cross- verification three times, calculating the conformity rate, reproduction rate, and F1 value mean.

Results Based on Bayesian optimization, CNNs had 75, 149, and 8 channels for Conv1–2, Conv3–5, and Conv6–8, respectively. The number of units for FC1–3 was 308, 476, and 783, respectively. The number of epochs was 44, the batch size was 3, and the AdaBelief learning rate was 2.51×10 -4 . The conformity rate, reproduction rate, and F1 value obtained from the five cross-verification tests are shown in Table 1. Table 1 The conformity rate, reproduction rate, F1 mean value, and standard deviation are shown in relation to each ROI. The diameter expresses the diameter of the circle expressed as a contour extraction error. ROI name Diameter (mm) Conformity rate ( ％ ) Reproduction rate ( ％ ) F1 value ( ％ ) Heart 8 99.8±0.2 99.0±0.6 99.4±0.3 12 99.9±0.1 98.5±0.9 99.2±0.4 Lungs 8 99.2±1.4 99.2±1.4 99.6±0.7 12 100.0±0.0 99.1±1.1 99.6±0.5

Conclusion By using CNN, the OAR’s circular contour error could be detected with an accuracy of 99% or more above the F1 value.

PO-1689 Sample pairing (mixup) as a data augmentation technique for deep medical image segmentation networks L.J. Isaksson 1 1 Istituto Europeo di Oncologia, Radiotherapy, Lund, Italy Purpose or Objective Researchers address the problem of generalizability of deep image processing networks mainly through extensive use of data augmentation techniques such as random flips, rotations, and deformations. A data augmentation technique called mixup, which constructs virtual training samples from convex combinations of