ESTRO 2021 Abstract Book

S1372

ESTRO 2021

inter-physician variability and the time cost of manual structure delineation have led to the investigation of auto-segmentation using deep learning methods. Our aim was to develop a convolutional neural network (CNN) for auto-segmentation of the gross tumor volume (GTV) and OAR for lung cancer and to assess its impact on clinical treatment plan evaluation. Materials and Methods We retrospectively selected the manual structure delineations of 47 patients treated with intensity-modulated radiotherapy IMRT for locally-advanced and 49 patients with stereotactic body radiation therapy (SBRT) for non-small cell lung cancer at our hospital between November 2018 and January 2020. Data augmentation and a U-Net architecture with pre-trained encoder weights were used to train a CNN for auto-segmentation of the following OAR: lungs, trachea, oesophagus, heart, spinal cord and left and right bronchus. For the SBRT patients only, a CNN was also trained for auto-segmentation of the GTV. Dice loss was used for CNN training and Dice score for CNN model assessment. In order to assess the clinical impact, a new treatment plan was optimized for 10 test patients based on the auto-segmented OARs. The resulting treatment plans were evaluated for the manual structures by comparing the dose-volume histograms and clinical goals. Results The average Dice score for the OAR was 0,809, ranging between 0,680 (oesophagus) and 0,974 (lungs). For the SBRT patients, the average Dice score for the GTV was 0,427. The auto-segmentation of the GTV was unsatisfactory, with additional bronchial structures being identified as target structure. For the clinical evaluation of the OAR, the mean absolute dose difference (averaged over all patients) between the manual and auto-segmented structures varied between 0,240 Gy (lungs) and 2,308 Gy (left bronchus). For each treatment plan based on the auto-segmented structures, the OAR clinical goals were also achieved for the manually delineated structures. Conclusion The results of the CNN were excellent for certain OAR, but its performance decreases for more complex OAR and for auto-segmentation of the GTV. The dosimetric impact of auto-segmentation on clinical treatment plan evaluation was on average not statistically significant, but may still be clinically relevant. Deep learning-based auto-segmentation can be used to alleviate the workload of the dosimetrists and/or physicians. Further refinement of the CNN should be performed to further improve the results and allow implementation in clinical practice. PO-1653 Measuring tissue thickness variation using Tomotherapy sinograms for H&N replanning M. Parisotto 1 , L. Reversi 1 , L. Alticozzi 1 , M. Valenti 1 , F. Fenu 2 , C. Di Carlo 2 , G. Mantello 2 1 Ospedali Riuniti di Ancona, Medical Physics, Ancona, Italy; 2 Ospedali Riuniti di Ancona, Internal Medicine, Ancona, Italy Purpose or Objective Head and Neck (H&N) radiotherapy (RT) with external photon beams is known to be a cause of weight loss of the irradiated tissue. This effect is a potential source of concern due to a lack of patient immobilization with an unfitting mask, lack of organ at risk dose sparing and target coverage. Helical treatments in Tomotherapy provide output sinograms whose intensity is affected by the attenuation of radiation passing through the patient. The purpose of this work is to evaluate a sinogram-based evaluation of inter-fraction thickness variation of patients that underwent H&N treatment with Tomotherapy, as a potential user-independent metric for driving re-planning decision. Materials and Methods A home-made Matlab software was delevoped to measure the water-equivalent thickness variation against the first fraction for patients who underwent H&N treatment with Tomotherapy. The software allowed to focus the analysis by filtering sinogram projections with beam direction, among Lat- Lat, AP-PA and "full" gantry rotation. Accuracy and repeatibility of measurement was evaluated on a pediatric CTDI phantom (PMMA of 10 cm diameter) used to mimic the neck of an adult. Different layers of bola of known thickness and density were arranged on the phantom top after the first fraction. Phantom alignment and target contours were set up such that only AP-PA projections passed through bola, Lat-Lat projections did not. Output sinograms of a set of five patients with initial Body Mass Index (BMI) of 27.7%±1.1% and consistent loss of weight at treatment end (DBMI = -11.8%±1.7%) were analyzed for thickness variation in LAT-LAT incidence. Dose prescription was 70Gy, 66Gy and 56.1-59.4Gy to surgical bed, high risk and low risk linphonodes, respectively, delivered in 33 fractions without replanning. We individuated four more patients, which have required replanning due to concern for diminishing of irradiated volume during H&N treatment with tomotherapy. Results Phantom thickness variation resulted within 5.7% of expected values and showed a good linearity with bolus thickness (R 2 = 0.997). Reproducibility evaluated in LAT-LAT incidence resulted within 0.1 cm.