PO57

Abstract

Purpose In the past years, a key improvement in the generation of treatment plans in high-dose-rate (HDR) brachytherapy comes from the development of multicriteria optimization (MCO) algorithms that generate thousands of pareto optimal plans within seconds. This brings a shift, from the objective of generating an acceptable plan to choosing the best plans out of thousands. Not only does the chosen plan depend on the planner, it also takes about 5-10 minutes to choose the preferred plan. The purpose of the present work is to speed up this process and to find a common ground for different specialists regarding the plan quality. Materials and Methods An AI algorithm based on the ResNet deep neural network architecture is developed to choose the best plan(s) from the generated plans. The algorithm classifies the plans, from the 3D dose distribution and anatomical structures, in 3 different classes, (1) violating hard (minimum) criteria, (2) respecting hard criteria and (3) respecting soft criteria, with every class being more stringent than the last one (increase in plan quality). The three classes are based on dosimetric criteria used at our institution for 15 Gy in a single fraction. For the classification, the more confident the model is that a plan belongs to class 3, the better is the plan. To mimic the behaviour of experts, visual-like criteria are implemented for the bladder, rectum and urethra. Visual criteria are defined as 100% and 125% isodose distance from the organ at risk. During training, the algorithm learns the link between the inputs (3D dose and anatomy) and outputs (visual-like and DVH's criteria). 850 previously treated prostate's cancer patients are used for the training and another set of 20 patients previously evaluated by two experts (clinical medical physicist) as part of an inter-observer MCO study are used for validation. For the training, 100 plans are generated for each patient using MCO and 27 000 plans are chosen at random to have the same quantity in each class. A NVIDIA GeForce RTX 3090 is used for training. Results The model takes 20 s to classify 2000 plans in order of preference (vs 5-10 mins for experts to rank 4 preferred plans). Currently, the training time is not optimized and it takes less than 2 days to train on the 27 000 plans with 75 epochs. For the 20 validation patients, 39.9 ± 20.2%, 46.4 ± 15.3% and 14.5 ± 21.9% of the plan are in class 1, 2 and 3 respectively. Table 1 shows the results obtained on 20 cases, each with 2000 plans; the mean and deviation are calculated based on the plan chosen by the model and by the experts. The table includes the best ranked and worst ranked plan of class 3. Looking at the best plan according to the model and comparing it with the plan chosen by the two experts show that the behaviour is similar. Out of the 40 chosen plans by the two experts, on 3 occasions our model ranked the same plan as the best plan. Looking more in depth, we find that the median ranking of the plans chosen by expert 1 and 2 is 71.5 and 136.5 respectively out of 2000. In one of the cases, there is no plan respecting the DVH criteria of class 3 and the result is suboptimal; the plan chosen by each expert does not respect only 1 of the criteria while the plan chosen by our model does not respect 3 criteria. This type of behaviour is undesirable and one of the next steps is to address this rare problem, where it is unfeasible to reach all criteria. Adding visual criteria restricted the number of plans which were considered for class number 3 from 16 500 (originally) to 9 000. Conclusions The approach is fast, adding negligible time to MCO planning, and preliminary results demonstrated the potential for clinical use. The approach is flexible with the possibility to adapt all criteria as desired. Future work will investigate model improvement, the non-inferiority of the best class 3 plan by the expert and methods to quickly restrict the number of navigated plans in order to obtain faster planning time. In the past years, a key improvement in the generation of treatment plans in high-dose-rate (HDR) brachytherapy comes from the development of multicriteria optimization (MCO) algorithms that generate thousands of pareto optimal plans within seconds. This brings a shift, from the objective of generating an acceptable plan to choosing the best plans out of thousands. Not only does the chosen plan depend on the planner, it also takes about 5-10 minutes to choose the preferred plan. The purpose of the present work is to speed up this process and to find a common ground for different specialists regarding the plan quality. An AI algorithm based on the ResNet deep neural network architecture is developed to choose the best plan(s) from the generated plans. The algorithm classifies the plans, from the 3D dose distribution and anatomical structures, in 3 different classes, (1) violating hard (minimum) criteria, (2) respecting hard criteria and (3) respecting soft criteria, with every class being more stringent than the last one (increase in plan quality). The three classes are based on dosimetric criteria used at our institution for 15 Gy in a single fraction. For the classification, the more confident the model is that a plan belongs to class 3, the better is the plan. To mimic the behaviour of experts, visual-like criteria are implemented for the bladder, rectum and urethra. Visual criteria are defined as 100% and 125% isodose distance from the organ at risk. During training, the algorithm learns the link between the inputs (3D dose and anatomy) and outputs (visual-like and DVH's criteria). 850 previously treated prostate's cancer patients are used for the training and another set of 20 patients previously evaluated by two experts (clinical medical physicist) as part of an inter-observer MCO study are used for validation. For the training, 100 plans are generated for each patient using MCO and 27 000 plans are chosen at random to have the same quantity in each class. A NVIDIA GeForce RTX 3090 is used for training. The model takes 20 s to classify 2000 plans in order of preference (vs 5-10 mins for experts to rank 4 preferred plans). Currently, the training time is not optimized and it takes less than 2 days to train on the 27 000 plans with 75 epochs. For the 20 validation patients, 39.9 ± 20.2%, 46.4 ± 15.3% and 14.5 ± 21.9% of the plan are in class 1, 2 and 3 respectively. Table 1 shows the results obtained on 20 cases, each with 2000 plans; the mean and deviation are calculated based on the plan chosen by the model and by the experts. The table includes the best ranked and worst ranked plan of class 3. Looking at the best plan according to the model and comparing it with the plan chosen by the two experts show that the behaviour is similar. Out of the 40 chosen plans by the two experts, on 3 occasions our model ranked the same plan as the best plan. Looking more in depth, we find that the median ranking of the plans chosen by expert 1 and 2 is 71.5 and 136.5 respectively out of 2000. In one of the cases, there is no plan respecting the DVH criteria of class 3 and the result is suboptimal; the plan chosen by each expert does not respect only 1 of the criteria while the plan chosen by our model does not respect 3 criteria. This type of behaviour is undesirable and one of the next steps is to address this rare problem, where it is unfeasible to reach all criteria. Adding visual criteria restricted the number of plans which were considered for class number 3 from 16 500 (originally) to 9 000. The approach is fast, adding negligible time to MCO planning, and preliminary results demonstrated the potential for clinical use. The approach is flexible with the possibility to adapt all criteria as desired. Future work will investigate model improvement, the non-inferiority of the best class 3 plan by the expert and methods to quickly restrict the number of navigated plans in order to obtain faster planning time.