ESTRO 38 Abstract book

S265 ESTRO 38

AUC 0.87 (95% CI 0.76-0.97) on the validation cohort. The best three models using radiomics with or without semantic features (semantic_dwi_post, semantic_dwi_adc_pre, t2_dwi_pre_post) were identified with performances on the development cohort with an AUC of 0.84 (95% CI 0.75-0.94), 0.85 (95% CI 0.75-0.98) and 0.83 (95% CI 0.70-0.95) respectively. Two models (semantic_dwi_post, t2_dwi_pre_post) validated well in the external patient cohort with an AUC of 0.86 (95% confidence interval 0.76 – 0.97) and 0.83 (95% confidence interval 0.70 – 0.95) respectively.

Radiomics machine learning approach for ADC with absolute rescaling and VOI delineation including tumor surroundings was useful for predicting locoregional recurrence of cervical cancer after definitive radiotherapy even using multi-center MRI data.

Intensive Rescaling

VOI

T2WI

DWI

ADC

Relative

VOI1 0.52 [0.390.65]

0.70 [0.59-0.80] 0.68 [0.57 0.79]

Relative

VOI2

0.66 [0.53-0.79]

0.73 [0.63-0.84] 0.5 [0.43-0.70]

Absolute

VOI1

N/A

N/A

0.67 [0.56-0.79]

Absolute

VOI2

N/A

N/A

0.79 [0.70-0.87]

OC-0510 MRI radiomics to predict tumour response in patients with locally advanced rectal cancer P. Bulens 1 , A. Couwenberg 2 , M. Intven 2 , A. Debucquoy 1 , V. Vandecaveye 3 , M. Philippens 2 , P. Mukherjee 4 , O. Gevaert 4 , K. Haustermans 1 1 University Hospital Gasthuisberg, Radiation Oncology, Leuven, Belgium ; 2 University Medical Center Utrecht, Radiation Oncology, Utrecht, The Netherlands; 3 University Hospital Gasthuisberg, Radiology, Leuven, Belgium; 4 Stanford University, Center for Biomedical Informatics Research, Stanford, USA Purpose or Objective To implement organ-sparing strategies into the multimodality treatment of patients with locally advanced rectal cancer (LARC), response prediction to select eligible patients is needed. In this research, we investigate the use of different multiparametric MRI-based radiomics models that predict (near-)complete response to chemoradiotherapy (CRT) in patients with LARC and compare their performances with the performance our previously developed and validated semantic model based on two volumetric and two ADC parameters. Material and Methods Radiomics models were developed in a cohort of 70 patients with LARC, prospectively recruited between 2012 and 2015. The external validation cohort consisted of 55 patients, recruited between 2008 and 2011. All patients were treated with CRT followed by surgery and underwent T2-weighted and diffusion-weighted imaging (DWI) before CRT and before surgery. The outcome measure for this study was (near-)complete pathological tumour response (ypT0-1N0). The tumour was segmented on T2-images and the ROI was transferred to DWI b800 images and ADC maps, after which radiomics features were extracted. Also, the two volumetric and two ADC parameters of the semantic model were calculated. Principal component analysis was used to linearly combine the radiomics and semantic features and regression analysis with LASSO was applied to develop the models. The best three models based on performance using receiver operating characteristic (ROC) and precision were selected for external validation and for comparison with the four-feature semantic model. Results 21/70 patients (30%) achieved ypT0-1N0 in the development cohort versus 13/55 patients (24%) in the validation cohort. The four-feature semantic model had a predicting performance of AUC 0.86 (95% confidence interval (CI) 0.76-0.95) on the development cohort and