ESTRO 2024 - Abstract Book

S4959

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Fig.2 The ROC curves for the test set using multivariate logistic regression models for predicting (a) grade ≥ 1 heamorrhage and (b) grade ≥ 2 haemorrhage.

Conclusion:

Combining radiomic features from planning CTs and daily MVCTs with dosimetric parameters, from both planned and accumulated doses, improves the predictive performance for rectal toxicities in prostate cancer patients after radiotherapy. Through this approach, there is potential to tailor treatments more precisely, minimising toxicities.

Keywords: radiomics, accumulated delivered dose, toxicity

204

Digital Poster

PET-CT DenseNet outperforms advanced DL models for outcome prediction of oropharyngeal cancer

Baoqiang Ma 1 , Jiapan Guo 1,2,3 , Lisanne V. van Dijk 1,4 , Johannes A. Langendijk 1 , Peter M.A. van Ooijen 1,2 , Stefan Both 1 , Nanna Maria Sijtsema 1 1 University Medical Center Groningen, Department of Radiation Oncology, Groningen, Netherlands. 2 University Medical Center Groningen, Machine Learning Lab, Data Science Center in Health (DASH), Groningen, Netherlands. 3 University of Groningen, Bernoulli Institute for Mathematics, Computer Science and Artificial Intelligence, Groningen, Netherlands. 4 University of Texas MD Anderson Cancer Center, Department of Radiation Oncology, Houston, USA

Purpose/Objective:

Treatment outcome prediction models could contribute to patient selection for personalized radiotherapy. In the HECKTOR 2022 challenge [1] many advanced deep learning models were presented for recurrence-free survival (RFS) prediction in head and neck patients based on PET- and CT- images. The goal of this work was to investigate whether a more common architecture such as DenseNet [2] with appropriate layer number and image-fusion settings could achieve comparable performance as those advanced models.

Material/Methods:

The HECKTOR 2022 dataset [1] includes 482 oropharyngeal cancer (OPC) patients from seven different centers. Another dataset consists of 400 OPC patients who received chemo(radiotherapy) at our center was also collected. In each patient, information on pretreatment CT, PET, manually generated GTV (Gross tumour volume) contours of primary tumors and lymph nodes, potential clinical predictors and RFS information was collected. The CT scans were only contrast enhancement in our dataset. The HECKTOR dataset was randomly split into a training set (n = 362) and an independent test set (n = 120) and our own dataset was used for external testing. Furthermore, the training set was divided into five folds that were stratified for event rate and tumour-stage for five-fold cross validation.