Abstract Book

S1088

ESTRO 37

EP-2000 Image-based data mining with continuous outcome variables W. Beasley 1 , A. Green 1 , A. McWilliam 2 , E. Vasquez Osorio 1 , M. Aznar 1 , M. Van Herk 1 1 The University of Manchester, Faculty of Biology- Medicine and Health, Manchester, United Kingdom 2 The Christie NHS Foundation Trust, Christie Medical Physics and Engineering, Manchester, United Kingdom Purpose or Objective Image-based data mining (IBDM) is a hypothesis- generating technique used to locate anatomy with dose- response relationships from previously treated radiotherapy patients without requiring prior region of interest delineation. Using deformable image registration, dose distributions are spatially normalised into a common frame of reference, and the dose at each voxel is related to a clinical outcome to locate organs or substructures showing a dose-response relationship. Previous IBDM applications used binary outcome variables (e.g. yes/no recurrence), but several clinical endpoints are continuous variables (e.g. amount of trismus after radiotherapy). We introduce here IBDM for continuous outcomes. Using simulated data, we compare this technique with traditional binary IBDM, for which continuous data are dichotomised with a threshold. Material and Methods The simulated dose images consisted of a set of 1000 128x128 images with a uniform background pixel value of 0.5 and a central 20x20 region in which the pixel value was set to increase from 0 to 1 across the set of images (Figure 1a). Gaussian noise was added to the dose images and each image was blurred with a Gaussian kernel, simulating incidental radiotherapy dose (Figure 1b). Each image in the set had a corresponding outcome label, which was a number that increased linearly between 0 and 100 across the image set. Gaussian noise was added to the outcome data to simulate a correlation <1 in the central region. For the simulation, 30 image-outcome pairs (i.e. a dose image and its corresponding outcome value) were randomly selected from the set of 1000 images. Binary IBDM using the t-statistic was performed using a range of thresholds for dichotomisation. Continuous IBDM used the Spearman correlation between each pixel value and the continuous outcome. In both techniques, permutation testing classified pixels with a statistically significant dose-response relationship. The classification of each pixel was compared to the true labelling and the sensitivity was calculated for each technique. This was repeated 100 times, calculating the average pixel classification, sensitivity and overall significance.

Results Continuous IBDM successfully identified the central region, whereas binary IBDM was only successful for thresholds close to the centre of the outcome distribution (Figure 2a). Continuous IBDM had a higher sensitivity and a lower statistical significance than binary IBDM for any threshold (Figure 2b).