Evaluation of the clinical use of a deep learning model for breast cancer treatment plan generation
Nienke Bakx,
The Netherlands
PD-0328
Abstract
Evaluation of the clinical use of a deep learning model for breast cancer treatment plan generation
Authors: Nienke Bakx1, Hanneke Bluemink1, Els Hagelaar1, Dave van Gruijthuijsen1, Jorien van der Leer1, Thérèse van Nunen1, Maurice van der Sangen1, Jacqueline Theuws1, Coen Hurkmans1
1Catharina Hospital, Radiation Oncology, Eindhoven, The Netherlands
Show Affiliations
Hide Affiliations
Purpose or Objective
The process of radiotherapy treatment planning is time-consuming due to manual and iterative steps. Besides, the outcome is dependent on the experience of the planner. The number of studies to automate this process with the help of artificial intelligence (AI) is increasing. The majority of these studies are of a retrospective nature, and only a limited number of AI models is actually implemented in the clinical routine. Recently, an AI model for dose prediction and automated treatment plan generation for left-sided whole breast radiotherapy was introduced in our clinic. In this study, the results for the first patients were monitored and evaluated.
Material and Methods
A deep learning (DL) dose prediction model, based on the U-net architecture, was developed by RaySearch (RaySearch Medical Laboratories AB), trained on an in-house collected dataset and implemented in our clinical workflow in May 2022, using RayStation TPS. Treatment plans were generated using the DL model. RTTs were allowed to manually adapt plans at their own discretion. The outcomes of the first 30 patients were evaluated by monitoring the changes made and quality was assessed based on the number of predefined clinical goals fulfilled for each patient.
Results
Figure 1 summarizes the outcomes of all 30 patients. 67% (20 out of 30) of the plans did fulfill all clinical goals without any adjustments. 7 of these plans were still adjusted, but it did not lead to any decrease in mean heart or lung dose. Most adjustments were made to decrease high dose areas and resulted in a mean decrease of 22 cGy of the D2% (range 1 – 53cGy), which is expected to not be clinically relevant. Of the 10 plans which did not directly fulfill all clinical goals, 6 could be easily adapted to fulfill these. For 4 of these patients the D98% was lower than the demanded 3805 cGy (ranging 3785-3799 cGy), whereas for the other 2 the D2% was higher than the allowed 4285 cGy (4296 and 4297 cGy). 2 out of 30 plans still did not fulfill clinical goals after manual intervention, although the mean heart dose was significantly decreased after manual adjustments for one patient from 296 cGy to 246 cGy.
Conclusion
The AI model successfully created a treatment plan in 67% of the cases and after small manual adjustments, 87% of the cases did fulfill all clinical goals. 50% of the manual adjustments did not decrease mean heart or lung dose. These results leave room for improvement of both training of the AI model to increase the number of cases fulfilling all clinical goals, as well as education of RTTs on clinically relevant adjustments.