ESTRO 2024 - Abstract Book

S3158

Physics - Autosegmentation

ESTRO 2024

Material/Methods:

Following institutional approval [Ethics approval number: 18/NW/0297], the population of patients with prostate cancer treated with low dose rate prostate brachytherapy (LDR PBT) between 2017 and 2022 was selected, with all patients having low or intermediate risk disease. All patients were imaged at the same site and with the same scanner/field strength (Siemens 1.5T) to remove additional sources of domain shift from the analysis. Of a total population of 241 patients, race information was available in 184 patients. White and Black patients represented the largest populations and were therefore selected for use in our experiments. To control for other possible confounders, we applied a matched pair design approach when selecting the cohorts such that each pair had similar prostate volumes with a tolerance of ±5 cm3 and similar age (±10 years). This resulted in 66 patients (50% White and 50% Black). We retrieved the diagnostic MR scans from the Picture Archiving and Communication System (PACS) for these patients and transferred them to the Varian Eclipse treatment planning system for anonymisation prior to model development and analysis. We employed the nnU-Net [4] state-of-the-art DL autocontouring framework to segment the prostate. All images used had ground truth prostate contours produced by an expert which were used for training and evaluating the nnU-Net model. We created five training datasets in which the proportions of the subjects varied by race: 0%/100% White/Black, 25%/75% White/Black, 50%/50% White/Black, 75%/25% White/Black and 100%/0% White/Black. We performed 4-fold cross-validation maintaining the matched pairs for each left-out fold, resulting in 24 subjects for training and 16 subjects (8 matched pairs) for testing per fold. To evaluate performance, we report the median and interquartile range of the dice similarity coefficient (DSC) across the entire dataset. P-values for Mann-Whitney U tests were computed to see if there was a significant difference between the performance of the 5 models on White and Black patients. The results that showed significant differences were further analysed using multivariate linear regression to assess the impact of different potential confounders. Selected variables analysed, in addition to race, were age, presenting PSA, prostate volume, slice thickness, BMI, and pixel spacing. The MATLAB [5] software was used to perform the statistical analysis

Results:

Table 1,2 summarises the results for each of the 5 autocontouring models broken down by the race of the subjects. The best results (DSC ≥ 0.9) for both races were obtained when an equal proportion from each race was used in the training set. The difference in DSCs was only significant for the model trained using 100% White subjects and 0% Black subjects, where White subjects had superior performance. Therefore, this model was further investigated for potential confounders using multivariate linear regression. Analysis of the pixel spacing showed that Black patient images had slightly better spatial resolution than the White subjects’ images but a multivariate linear regression showed that this difference was not significant at a significance level of 0.05. The only variable that was statistically significant was race.

Table 1