Feasibility of image-based data mining in breast radiotherapy

Abstract

Purpose or Objective Breast pain following cancer treatment reduces the patient’s quality of life. Although the correlation between pain and radiation hotspot has been established in previous studies, the impact of the location of the hotspot has not been identified. Voxel-wise image-based data mining (IBDM) can identify specific dose patterns correlated with radiotherapy toxicity and identify sensitive sub-regions. However, IBDM has not been applied to breast radiotherapy as a consequence of variations in volumes (breast size) and set-up positions in this patient population. We aimed to demonstrate the feasibility of applying IBDM to breast cancer patients treated with supine post-lumpectomy radiotherapy in a large multi-centre study. Materials and Methods IBDM was applied to 177 patients from 8 centres in the REQUITE study (www.requite.eu). All patients were treated supine with different arm positions; both arms up (n=118), left arm up (n=32), or right arm up (n=27). Planning dose distributions were normalised to a single reference patient using NiftyReg deformable image registration (DIR). The region of interest was the breast +/-5 cm in the superior-inferior direction. DIR performance was assessed using the normalised correlation coefficient (NCC). The dose in each voxel was converted into EQD2 using α/β=1.7 for moderate or marked normal tissue toxicity in the breast or chest wall. For this feasibility study, the outcome of interest was patient-reported pain at 1-year post-radiotherapy (any pain vs no pain). The region associated with pain was identified using voxel-wise t-test and permutation testing (n=1000). The correlation between breast pain and mean and max EQD2 in the identified region, as well as other clinical variables were investigated using univariable and multivariable logistic regression analysis (SPSS v.25). Results Patient characteristics are summarised in Table 1. DIR accuracy on the breast was acceptable (NCC=0.92 (IQR 0.14), where 1 is the ideal value), but was considerably affected by the difference in arm positions (Mann-Whitney test between both arms up and ipsilateral arm up only, p=0.05). IBDM identified a sensitive region which correlated with breast pain in the upper part of the treated breast (Figure 1.A-B). The mean and max EQD2 within the identified region, fractionation, and breast cup size were significant in univariable analysis (Figure 1.C). However, only fractionation and breast cup size remained significant on multivariable analysis. Conclusion IBDM is feasible in breast radiotherapy and the accuracy of DIR was found acceptable. Nevertheless, larger cohorts of patients are needed to clarify the existence of a sensitive sub-region related to breast pain and/or other side effects. Further work will investigate increasing the power of the analysis by mirroring the patients’ anatomy to overlay left/right sided dose distribution, and including patients treated in the prone position.