(P 229) ON THE VARIABILITY IN THE ANALYSIS OF APPARENT DIFFUSION COEFFICIENT ESTIMATES FOR MULTICENTER DIFFUSION MRI DATA OF RHABDOMYOSARCOMA PATIENTS

Objective: Pediatric rhabdomyosarcoma (RMS) is a common and aggressive type of soft tissue sarcoma that requires multimodal therapy and can be found in different parts of the body. In this study, retrospective data from a multicenter longitudinal study (EpSSG RMS-2005 study) is used to evaluate the diagnostic potential of diffusion MRI in RMS. As multicenter studies become increasingly more common in clinical trials, it is crucial to acknowledge that apparent diffusion coefficient (ADC) values may be influenced by various factors such as acquisition parameters, scanner manufacturers, pre and post processing, type of segmentation and pipeline used to create the ADC maps. Our goal is to compare ADC map generated by different scanners to ADC maps computed by a standardized in-house pipeline from diffusion weighted images (DWI), to determine whether these factors might contribute to a potential decrease in sensitivity in measurement across multiple centers.

Methods: In our study we included 144 sets of images from 67 pediatric patients. Each set has: an ADC map automatically generated from the scanner, an ADC map computed by our in-house pipeline using existing DWI data, and a binary mask that delineates the non-necrotic regions of the tumor. In both ADC maps, outliers from noise were removed using the 1.5 interquartile range rule. Among the patients, longitudinal data was available for 36 individuals, consisting of one scan at diagnosis and another after induction chemotherapy (follow up). The study aims to compare ADC values obtained from the scanner with those calculated from DWI within the tumor across all images (I), and with focus on longitudinal data (II), using relative percent difference between images and difference of ADC values for each voxel within the mask. Paired t-test was used for statistical analysis.

Results: (I) In each of the 144 sets of images (few examples are shown in Fig. 1), the average voxel-wise relative percentile difference between ADC from scanner and DWI within the tumor is 5.94 ± 27.94 % (range -8.30, +200.00 %, median value 0.89%). This is linked to a median voxel-wise difference in ADC values (ADC from DWI – ADC from scanner) of 0.04 ± 0.18 ·10-3 mm2/s (range -0.15, +1.58 ·10-3 mm2/s, median value 0.01·10-3 mm2/s). No statistical difference in median relative difference has been found for different manufacturers (GE, Philips, Siemens).

(II) Fig. 2 depicts an example of scan at diagnosis and follow up, and relative histograms of ADC from scanner and DWI. As shown in Fig. 3, the median ADC values from the scanner at diagnosis are 1.16 ± 0.32 ·10-3 mm2/s, and at follow-up are 1.51 ± 0.36 ·10-3 mm2/s. On the other hand, median ADC values from DWI at diagnosis are 1.18 ± 0.31 ·10-3 mm2/s and at follow-up are 1.55 ± 0.35 ·10-3 mm2/s. Taking into consideration the median values of the segmented areas, statistically significant differences have been found between diagnosis and follow up, but not between ADC from the scanner and DWI at the same time point.

Conclusion: Our results show that the relative difference within the tumor between ADC from scanner and DWI can vary substantially. As shown in Fig. 1, difference in ADC values might be caused by the different pipeline used to generate the image, as well as preprocessing steps on the ADC map from scanner. Even if differences in median ADC values from scanner and DWI do not have a statistical difference in this cohort, we see how ADC values at diagnosis have consistently lower values compared to the ones at follow up (Fig. 2, Fig. 3). It is important to be aware of these differences, as they can potentially create another layer of variability in a multicenter study. Moreover, ADC maps generated from different scanners can undergo different pre and postprocessing steps, of which the user might not be aware of. A standardized pipeline for generating ADC maps across all images in the dataset would mitigate unwanted variability and enhance reliability and comparability across the dataset.