Limitations of Automated Segmentation tools for Breast Tissue In Young Women Treated with Radiotherapy to the Chest

The performance of commercially available segmentation tools (deep learning and atlas-based) were assessed for breast contouring of young lymphoma patients on CT. Dice similarity coefficient, mean distance to agreement and Hausdorff distance were used to analyse performance. Deep learning segmentation performed better on more patients (6/10) but atlas-based segmentation performed best on 2/10. The variation of breast densities and arm positions likely affects auto-contouring performance in young lymphoma patients, as atlas libraries struggle to encompass the wide variation, and deep learning training data is typically of older breast cancer patients.