Automated deep learning segmentation and planning for left-sided breast radiotherapy with minimised adaptations based on dose, TCP and NTCP criteria

Background and Purpose: Deep learning segmentation (DLS) and planning (DLP) models in radiotherapy treatment planning increase consistency and reduce planning time. However, many DL segmentations and plans are manually adjusted. The clinical relevance of these adjustments remains underexplored. This study assesses the clinical relevance of adjustments made to DLS and DLP and proposes a workflow that could optimise treatment planning time while remaining clinically similar. Materials and Methods: A retrospective analysis of 101 clinically approved structures and treatment plans (CS-CP) for left-sided node-negative breast cancer was conducted, comparing them with three alternative plans varying in DL influence: unadjusted DLS and DLP (DLS-DLP), clinically approved structures with DLP (CS-DLP), and clinically approved targets with DLS organs-at-risk, referred to as proposed structures, and DLP (PS-DLP). Plans were evaluated based on clinical goals, tumour control probability, and normal tissue complication probability. A new workflow, with four planning routes, was designed to assess time savings. Results: 88 % of CS-CP plans met all clinical goals, compared to 43 % (DLS-DLP), 59 % (CS-DLP) and 60 % (PS-DLP). Statistically significant differences were found for most evaluation criteria, with only the D98 and tumour control probability showing clinically relevant differences. The new workflow reduced active planning time by 45 min and overall workflow time by 12 h. Conclusion: The clinical relevance of DLS and DLP adjustments was assessed, revealing that plans generated by a fully DL workflow (DLS-DLP) were clinically acceptable for 43 % of patients. The proposed workflow optimised planning time, enhancing efficiency.