Artificial Intelligence Algorithms to Automate the Total Marrow Irradiation: The AuToMI Project

Total marrow and (lymphoid) irradiation (TMI/TMLI) is a radiotherapy (RT) treatment that selectively irradiates the entire bone marrow and lymphatic system. TMI/TMLI is used in conditioning regimens before hematopoietic stem-cell transplantation for acute leukemia and multiple myeloma. Compared with traditional total body irradiation (TBI), where the entire body of the patient is irradiated, TMI/TMLI offers a targeted approach with reduced risk of toxicity and improved efficacy. This treatment can be delivered with volumetric modulated arc therapy (VMAT) and many centers worldwide have the potential to implement TMI/TMLI using this technique. However, technological and knowledge gaps currently limit its widespread introduction as a modern alternative to TBI. This thesis was divided into two main parts. The first part (Papers I-IV) focused on streamlining the workflow of VMAT-based TMI/TMLI and improving its accessibility for centers which intend to adopt this treatment in the near future. The second part (Papers V-VII) explored methods to reduce patient-specific quality assurance (PSQA) workload using the entire database of VMAT plans delivered at IRCCS Humanitas Research Hospital. In Paper I, a deep learning-based approach was developed for automatically generating the field geometry of TMI/TMLI for the upper body, providing results on par with an experienced planner. In Paper II, the dosimetric quality and complexity of the treatment plans were analyzed. In particular, correlations between complexity and PSQA outcome were investigated. Then, in Paper III, an automated planning tool was developed for creating a robust plan-to-plan field junction. The software reduced the time required to optimize TMI/TMLI treatments by up to two days. Finally, in Paper IV, patient positioning was optimized to reduce unwanted movements. The optimal patient positioning was identified as having the arms alongside the body and the feet immobilized with a cushion. In Paper V, a machine learning (ML) model was developed to predict the PSQA outcome for VMAT plans. The model was validated in a multicentric scenario and produced conservative estimates. In Paper VI, the Lean Six Sigma methodology and ML model were implemented in clinical practice to continuously supervise the RT optimization process and enable a targeted PSQA approach. This method allowed to direct the attention and resources of the clinical staff to rare events. Finally, Paper VII explored replanning strategies to improve deliverability while maintaining dosimetric quality of plans identified at risk by the ML model. Automating these procedures could reduce the monthly workload by over 10 hours.