(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Mathematical models have been proposed to capture the correlation between plan dosimetric parameters and patient anatomical variations as baseline for plan quality evaluation. However, despite recent development in machine learning methods, there is no systematic study on how to handle the heterogeneity in multi-institution data, e.g. patient anatomy, physician preference, contouring variability, treatment planner experience. The purpose of this study is to develop a framework to evaluate the quality of treatment plans in a multi-institutional setting.
Methods: We have developed a workflow to first group 395 prostate cancer treatment plans from 38 institutions into clusters automatically by hierarchical clustering. Separate models are then trained for each cluster. The similarity between different institutions data was defined as the improvement (or degradation) of the model prediction accuracy trained on the combined data from these institutions compared with the models trained on each individual one. The clusters continue growing pair-wise until the model prediction accuracy degrades significantly. The models looked at anatomical factors such as distance-to-target-histogram and other volumetric factors and utilized a step-wise regression method for feature selection. The model performances are compared with the global models which were trained by all the institution data together.
Results: Three relevant QUANTEC dosimetric indices [D15, D25, Dmean] were predicted for bladder, rectum and bowel using the models. The determination coefficients (R2) by the global model were [0.50, 0.62, 0.69] for each dosimetrics in bladder, [0.46, 0.46, 0.41] in rectum, [0.56, 0.53, 0.51] in bowel, respectively. The clustered model divided the 38 institutions into 5 clusters, with 16, 7, 6, 5, 4 institutions in each cluster respectively. The clustered model improved the prediction accuracy for rectum to [0.64, 0.58, 0.65].
Conclusion: proposed plan evaluation framework can capture plan quality variations in the presence of patient anatomical variation and large data heterogeneity.
Keywords
Taxonomy
TH- Dataset Analysis/Biomathematics: Machine learning techniques
Contact Email
