ESTRO 2022 - Abstract Book

S1568

Abstract book

ESTRO 2022

Results For radiomics, the best pipeline was formed by the combination of feature agglomeration method and bagging classifier, which achieved an area under the receiver operating characteristic curve (AUC) of 0.683 and 0.652 on the training and test set, respectively. In general, deep learning approaches performed slightly worse than radiomics, both in terms of AUC and of difference between training and test performance. For the deep feature approach, the AUC of the best model was 0.692 and 0.631 on the training and test set, respectively, while for the CNN approach the greatest AUC was 0.692 and 0.644, respectively. The reduced predictive capabilities of deep learning methods could be ascribed to the bidimensional restriction of the analysis - necessary to reduce model complexity and increase the dataset size – which entails a poorer description of the tumour masses. Conclusion Using the public LUNG1 dataset, we extensively investigated and compared the performance of several classification pipelines based either on radiomic or on deep learning approaches for the prediction of the 2-year OS in NSCLC. The results of the best pipeline were comparable among the approaches and in line with previous works, demonstrating that these techniques are not able to extract additional information. However, we expect that the application of 3D deep learning models, in this work limited by the reduced size of the dataset, would increase the performance. 1 University of Gothenburg, Sahlgrenska Academy, Institute of Clinical Sciences, Department of Medical Radiation Sciences, Gothenburg, Sweden; 2 Sahlgrenska University Hospital, Medical Physics and Biomedical Engineering, Department of Therapeutic Radiation Physics, Gothenburg, Sweden Purpose or Objective NTCP models are essential in RT as the basis for population-based planning objectives in treatment planning. The models result from analyses of clinical data, such as logistic regressions to estimate the relationship between the dose to an OAR and the risk of an outcome. However, the dose used to create a model, usually retrieved from a delineated volume in a treatment planning system, is just an estimation of the delivered dose to the organ. This estimation includes errors caused by for example calculation and delineation errors. The purpose of this study was to simulate how errors in the dose values used in dose-response modelling affects the resulting NTCP model. Materials and Methods We simulated the dose-response modelling procedure by creating reference populations from a defined distribution of dose values with an assumed dose-response relationship, and introducing dose errors repeatedly. Dose values within each reference population were randomly generated from a normal distribution, 16±4 Gy (mean±1 SD) truncated at ±3 SD. This corresponds to the distribution of mean lung doses when treating lung cancer with conventional RT at our hospital. Four general scenarios were considered: population risks of either 10 or 25% in combination with study sizes of either 100 or 300 patients. Each patient within the reference populations was assigned an endpoint based on an NTCP model where the dose-response parameter γ 50 was set to 1.0 and the D 50 was calculated to match the population risk in each scenario. If a randomly generated number between 0 and 100% was less than or equal to the calculated NTCP, an endpoint of 1 was given, otherwise 0. We randomly created 50 reference populations. For all reference populations, we estimated their reference NTCP models using logistic regression and the 95% CIs with bootstrapping. For each reference population in each scenario, we repeatedly (1000 times) introduced random relative dose errors to the reference doses. The errors were randomly sampled from normal distributions (0±5% or 0±10%) truncated at ±3 SD. New NTCP models were estimated using the doses with introduced errors. Results The larger dose error resulted in larger systematic over- and underestimations of the NTCP for doses below and above 16 Gy, respectively, see Figure 1 and 2. PO-1764 Impact of dose errors on dose-response modelling L. Mövik 1 , A. Bäck 2,1 , N. Pettersson 2,1