Lesser of Two Evils Improves Learning in the Context of Cortical Thickness Estimation Models - Choose Wisely

Cortical thickness (CTh) is an important biomarker commonly used in clinical studies for a range of neurodegenerative and neurological conditions. In such studies, CTh estimation software packages are employed to estimate CTh from T1-weighted (T1-w) brain MRI scans. Since commonly used software packages (e.g. FreeSurfer) are time-consuming, the fast-inference Machine Learning (ML) CTh estimation solutions have gained much popularity. Recently, several ML regression-based solutions offering morphological properties (CTh, volume and curvature) estimation have emerged but typically achieved lower accuracy compared to mainstream alternatives. One of the reasons for such performance of the ML-based CTh estimation models is the inaccurate automatic labels typically used for their training. In this paper, we investigate the impact of automatic labels selection on the performance of the current state-of-the-art ML regression-based CTh estimation method - HerstonNet. We train two models on pairs of brain MRIs and FreeSurfer/DL+DiReCT automatic CTh measurements to investigate the benefits of using DL+DiReCT instead of, the more frequently used, FreeSurfer CTh measurements on the learning capability of a modified version of HerstonNet. Then, we evaluate the performance of the two trained models on three test sets with scans coming from four publicly available datasets. We showthatHerstonNet trained on DL+DiReCT labels overall achieves a 13.3% higher Intraclass Correlation Coefficient (ICC) on a test set composed of ADNI and AIBL scans, 19.4% on OASIS-3 and 17.1% on SIMON dataset compared to the same model trained on FreeSurfer derived measurements. The results suggest that DL+DiReCT provides automatic labels more suitable for CTh estimation model training than FreeSurfer.