ESTRO 2023 - Abstract Book

S1881

Digital Posters

ESTRO 2023

greater than 0.8 for all models. CART analysis classified patients with IM,T0 ≥ 99 to be associated with RTOG ≥ 2 toxicity; subsequently, PTV1 and PTV2 played a significant role in increasing the classification rate. The CART model provided a very high diagnostic performance of AUC=0.959. Conclusion Spectrophotometry is an objective and reliable tool able to assess radiation induced skin tissue injury. Using a machine learning approach, we were able to predict grade RTOG ≥ 2 skin toxicity in patients undergoing breast RT. This approach may prove useful for treatment management aiming to improve patient quality of life.

PO-2098 Deep learning post-radiotherapy recurrence prediction in head and neck cancer from planning PET/CT

D. Kutnar 1,2

1 University of Copenhagen, Computer Science, Copenhagen, Denmark; 2 Rigshospitalet, Department of Oncology, Copenhagen , Denmark Purpose or Objective Prediction of post-RT loco-regional recurrence risk and position (LRR) could potentially be used to guide personalized radiotherapy. Convolutional neural networks (CNNs) have demonstrated promise across a wide range of medical domains, yet to the best of our knowledge, no prior work has investigated the potential of CNNs to segment recurrence volume from imaging at baseline. Therefore we investigated the extent to which a CNN is able to segment LRR from pre-treatment PET/CT scans in HNSCC patients. Materials and Methods This study employed 37 patients who had undergone primary RT for OPSCC at Rigshospitalet, Denmark. All patients developed biopsy-verified isolated LRR in the median 1.3 years after treatment completion. Five experienced oncologists contoured the relapse volumes on recurrence CT scans using Eclipse (Varian). For model development, the corresponding pre-treatment PET/CT, GTV and contoured relapse for each of the patients were randomly divided into training (n=23), validation (n=7) and test (n=7) datasets. We compared a trained CNN to an SUVmax thresholding method (50% of SUVmax) and to using the GTV contour directly as the recurrence prediction. All comparisons are on the test set only. We computed the number of relapse origin points included in the predicted region for each method. Here the origin is defined as the central point of each manually contoured relapse volume. We performed a paired t-test to compare the methods on the test set in terms of predicted volume. Results The SUVMax threshold method included 5 out of the 7 relapse origin points within a volume of median 34 cc. Both the GTV Contour and CNN segmentations included the relapse origin 6 out of 7 times with median volumes of 39 and 20 cc respectively.

TABLE I: Methods comparison between seven patients (P1 - P7) who have developed LRR included in the test dataset. Bold font indicates best result among methods. Conclusion CNN segmentation included the same number of relapse origin points whilst providing a significantly smaller predicted relapse volume than GTV. This suggests that a CNN defined volume may be helpful in defining a subvolume of GTV to boost in high risk patients. Another advantage of the CNN output is that an adjustable threshold can be applied, allowing the method to be tuned based on the clinically desired balance of sensitivity and specificity or to provide a recipe for dose painting by numbers.