Clinical Evaluation of a Deep Learning Model for Segmentation of Nodal Clinical Target Volumes in Breast Cancer Radiotherapy

Precise segmentation of clinical target volumes (CTV) in breast cancer is indispensable for state-of-the art radiotherapy. Despite international guidelines, significant intra- and interobserver variability exists, potentially negatively impacting treatment outcomes. The aim of this study is to evaluate accuracy and efficiency of segmentation of nodal CTVs in planning CT images of breast cancer patients performed by a 3D convolutional neural network (CNN) compared to the manual process. An expert radiation oncologist (RO) segmented 6 different nodal CTVs (levels IV through I, Rotter’s space and the Internal Mammary Nodes) according to international guidelines in 150 breast cancer patients. This data was used to create, train and cross-validate the CNN. The network's performance was further clinically evaluated using a test set of 20 patients. In addition to the expert RO, a sample of 5 resident ROs active in daily clinical practice each performed manual segmentation of 4 patients in the test set and were blinded for the CTVs generated by the CNN. Quantitative analysis of CTV segmentation by the CNN using Dice Similarity Coefficient (DSC) was performed on the test set, using CTVs generated by the expert RO as ground truth. Qualitative analysis for accuracy was performed using a predefined checklist with 34 possible major and 42 possible minor guideline deviations (i.e. errors against anatomical boundaries) for the 6 CTVs combined. Results of the manual process were then compared to the results of the output generated by the CNN. Efficiency was assessed by comparing the time needed to correct CTVs generated by the CNN and time needed for manual segmentation. Mean DSC over all nodal levels generated by the CNN was 0.73 with a standard deviation of 0.07. Qualitative scoring of accuracy of the CNN output showed an absolute decrease of 8.35% in major guideline deviations (23.15% to 14.80%) and an absolute decrease of 14.48% in minor guideline deviations (28.78% to 14.30%) when compared to manually generated volumes. The majority (71%) of guideline deviations in the test set of the CNN consisted of errors in cranial or caudal margins. For the CNN output, the mean correction time was 11 minutes. This was 24 minutes shorter in comparison to the mean time required for manual segmentation (35 minutes). The CNN outperformed ROs for segmentation of nodal CTVs with regard to major and minor deviations from guidelines. Furthermore, the time required to acquire these CTVs decreased significantly. The majority of remaining guideline deviations in target volumes predicted by the CNN consists of errors in the cranial and caudal margins. This study is the first to evaluate the role of deep learning in nodal CTV definition in breast cancer radiotherapy, proving its potential to further increase uniformity and efficacy in the segmentation process.