Borrowing strength from adults: Transferability of AI algorithms for paediatric brain and tumour segmentation

AI brain tumour segmentation and brain extraction algorithms promise better diagnostic and follow-up of brain tumours in adults. The development of such tools for paediatric populations is restricted by limited training data but careful adaption of adult algorithms to paediatric population might be a solution. Here, we aim exploring the transferability of algorithms for brain (HD-BET) and tumour segmentation (HD-GLIOMA) in adults to paediatric imaging studies.

In a retrospective cohort, we compared automated segmentation with expert masks. We used the dice coefficient for evaluating the similarity and multivariate regressions for the influence of covariates. We explored the feasibility of automatic tumor classification based on diffusion data.

In 42 patients (mean age 7 years, 9 below 2 years, 26 males), segmentation was excellent for brain extraction (mean dice 0.99, range 0.85-1), moderate for segmentation of contrast-enhancing tumours (mean dice 0.67, range 0-1), and weak for non-enhancing T2-signal abnormalities (mean dice 0.41). Precision was better for enhancing tumour parts (p < 0.001) and for malignant histology (p = 0.006 and p = 0.012) but independent from myelinisation as indicated by the age (p = 0.472). Automated tumour grading based on mean diffusivity (MD) values from automated masks was good (AUC = 0.86) but tended to be less accurate than MD values from expert masks (AUC = 1, p = 0.208).

HD-BET provides a reliable extraction of the paediatric brain. HD-GLIOMA works moderately for contrast-enhancing tumours parts. Without optimization, brain tumor AI algorithms trained on adults and used on paediatric patients may yield acceptable results depending on the clinical scenario.