Session Item

Radiotherapy research increasingly focuses on managing long-term morbidity and mortality of curative treatments. Such research relies on knowledge of the relation between radiation exposures and their biologic effects. Critically, this relation is often uncertain or unknown. Image-based data mining (IBDM) is a voxel-based method for analyzing dose response that has shown promise in studies of adults with cancer. Age-related anatomic variations, however, complicate its applicability to children. We recently successfully applied IBDM to children with simplified simulated radiation treatments. In this study, we tested its efficacy using simulated but clinically realistic dose distributions.

We used CT images from 167 children (age 10 months to 20 y) who previously received radiotherapy for primary brain tumors. We randomly divided the cohort into simulated “effect” and “no effect” groups and modified their clinical dose distributions by introducing a systematic dose discrepancy between the brainstems of patients in each group. The IBDM method comprises two steps: deformable image registration (DIR) to a reference anatomy and voxelated dose comparison. Based on previous results, we selected the CT dataset from the patient with the median brain volume as the reference anatomy. We quantified the accuracy of DIR via contour-distance and center-of-mass analyses in 5 routinely delineated brain structures. Dose comparisons comprised permutation tests with 1000 permutations. We quantified the voxel-wise accuracy with sensitivity, positive predictive value (PPV), and dice similarity coefficient (DSC) comparing the reference brainstem to the volume in which IBDM identified a dose discrepancy. We performed these tests at three simulated “effect” rates, a baseline of 50% and two rates representative of common side effects of cranial irradiation in children: 20% and 10%.

DIR accuracy was < 2 mm (average contour distance) and < 3 mm (center of mass). Permutation tests correctly identified the difference between dose distributions of the “effect” and “no effect” groups (p < 0.01) at all three simulated effect rates (Table). IBDM consistently recognized the dose differences in the brainstem with a minimum sensitivity of 0.98, but over-estimated the sensitive region, likely because of implicit correlations in the dose distributions (Table, Figure). PPV and DSC improved with decreasing effect rate as the overestimate of the sensitive region was reduced, reaching maxima of 0.25 and 0.4, respectively, at the 10% simulated effect rate (Table, Figure).

This work shows it is feasible to perform IBDM for pediatric data despite large anatomic variations among patients. This approach enables an unbiased exploration of the spatial dose-effect relation in children. A deeper understanding of this relation can inform treatment planning and survivorship care to avoid treatment-related morbidity and mortality and improve long-term quality of life of childhood cancer survivors.