ESTRO 2024 - Abstract Book

S4574

Physics - Machine learning models and clinical applications

ESTRO 2024

Across all quality metrics, plans generated using the KBP pipeline performed at least as well as or better than the respective clinical plans. The average conformity and gradient of IO plans were 0.737 ± 0.158 and 3.356 ± 1.030, respectively, compared to 0.713 ± 0.124 and 3.452 ± 1.123 for the clinical plans. IO plans also outperformed DM plans for five of the six quality metrics. Additionally, plans generated using the IO pipeline had an average treatment time comparable to clinical plans. Table 1 presents the average quality metrics and treatment time for the 27 test patients along with their standard deviation.

Clinical

Inverse Optimization

Dose Mimicking

Coverage index

0.950 ± 0.032

0.945 ± 0.087

0.921 ± 0.122

Selectivity index

0.751 ± 0.134

0.784 ± 0.166

0.694 ± 0.219

Conformity index

0.713 ± 0.124

0.737 ± 0.158

0.641 ± 0.207

Gradient index

3.452 ± 1.123

3.356 ± 1.030

3.727 ± 1.279

Quality of coverage

0.836 ± 0.077

0.879 ± 0.098

0.818 ± 0.153

Homogeneity index

1.778 ± 0.266

1.830 ± 0.378

1.663 ± 0.185

Treatment time (min)

35.08 ± 34.66

32.92 ± 28.76

124.6 ± 100.0

Table 1: Average and standard deviation in coverage index, selectivity index, conformity index, gradient index, quality of coverage, and homogeneity index for all plans. Overall treatment time is also listed in minutes.

The average time required to convert a patient’s data into a deliverable plan using the pipeline was 5 minutes 43 seconds. This time included data modification to generate a prediction using the trained model, solving the inverse optimization and inverse planning models, and solving an additional optimization problem to generate the appropriate shots. The overall calculation time varied depending on target complexity. Compared to clinical plans, KBP plans utilized block sectors significantly more frequently and 4 mm collimators significantly less frequently. Additionally, KBP plans favor using multiple shots per isocenter in contrast to manual clinical plans, which are based on one shot per isocenter.

Conclusion:

Plans resulting from an IO KBP pipeline consistently match or surpass the quality of manual plans. The results demonstrate the potential for the usage of KBP to generate GK treatment plans with minimal human intervention

Keywords: Gamma Knife, optimization, knowledge-base planning

References: