Abstract Background: Breast cancer is the most common cancer in women, and radiotherapy is an integral part of treatment for most patients. However, as with any treatment, radiotherapy is associated with toxicity, such as fibrosis, ulceration, wet or dry desquamation (Skin peeling), erythema (skin redness), lymphedema (swelling of lymph nodes) and others. The REQUITE study is a large, multi-institution study, dedicated to furthering understanding of radiation-induced toxicity and facilitating improvements in future treatments. In this project, we provide a first analysis of radiotherapy images in a large cohort of patients to investigate how calculated volumetric densities correlate to age and volume. Materials and Methods: 1000 CT scans and associated dose distributions for breast radiotherapy patients were available from the REQUITE dataset. The methodology was first established in asubset of 114 patients (cohort 1) before being applied to the whole dataset (cohort 2). In this work, we focused on the following objectives: 1) ensuring consistent delineations of the breast volumes both for the treated breast (ipsilateral) and the healthy breast (contralateral) and 2) analysing the breast density based on CT images. Automatic breast contours were created using two different atlas-based automatic contouring software, ADMIRE (Advanced Medical Image Registration Engine by Elekta) and Mirada DBx (atlas-based image segmentation tool by Mirada Medical Ltd). These were compared to the manual experts contour available in 96 patients (ipsilateral supine positions) and 96 patients (contralateral supine positions) in cohort 1. Performance was assessed using SÃ¸rensenâ€“Dice coefficient (statistic used to measure contour similarity). Once validated, automatic volumes were available. The density of the breast was then investigated by thresholding voxels within the volumes, using threshold Otsu and pixel intensity ranges based on Hounsfield units. The range of Hounsfield units were -200 to -100 for fatty tissue, and -99 t0 +100 for fibro-glandular tissue. The volumetric breast density (VBD) was defined as volume of fibro glandular tissue / (volume of fibro glandular tissue + volume of fatty tissue). This metric was chosen to ensure that we are able to validate the breast densities in future using mammographic images. A sensitivity analysis was performed to verify whether the calculation of VBD was affected by the choice of breast contour (whether generated by Mirada, ADMIRE, or manually). In addition, we investigated the correlation between VBD, volume of the breast, and age of the patient using statistical analyses. VBD values were compared between ipsilateral and contralateral breast contours. Results: Both ADMIRE and Mirada resulted in equally high Dice coefficients (ADMIRE range of dices were 0.72 to 0.97 with median 0.91. Mirada dices ranged from 0.75 to 0.94 with median 0.90) compared to expert contours in cohort 1. However, ADMIRE was easier to use on a large dataset, and was chosen for application to cohort 2. The breast density / VBD ranged from 0.17 to 0.91 (with median 0.40) in cohort 1, and 0.096 to 0.99 (with median 0.43) in cohort 2. The sensitivity analysis performed on cohort 1 revealed that the density was not significantly affected by the choice of contours). From analyses of ipsilateral and contralateral breast contours, ipsilateral breasts tend to be denser than contralateral breasts. Breast density was negatively associated with breast volumes (Spearman: R=-0.5, p-value < 2.2e-16) and age (Spearman: R=-0.24, P-value = 4.6e-10) in cohort 2. Same test was repeated for breast volumes 500 cc or higher (Age vs Density: R=-0.1, P-value = 0.029 , Volume vs Density: R=-0.19, P-value = 4.4e-05) Conclusion: ADMIRE was reliable in processing cohort 2 due to the automatic scripting feature which was not available in Mirada DBx. However, main findings from our studies include reliably verifying that it is possible to contour both the ipsilateral and contralateral breasts with automatic delineation methods. We were able to develop an approach to calculate breast densities from CT scans, which will be validated later. We also found that breast density may not entirely be explained from variables such as patientsâ€™ age or breast volume. As a result, we are able to justify investigating our CT-estimated densities as variables in a risk prediction model.
- breast cancer
- deep learning
- big data
- breast density
BIG DATA FOR BETTER BREAST CANCER TREATMENT
Akuoko, D. (Author). 1 Aug 2021
Student thesis: Master of Philosophy