This present study only examined one modality of MRI, that is structural MRI. The results from this study might therefore not extend to data from different modalities, such as functional MRI (fMRI) or diffusion weighted imaging. Especially in the case of constructing connectome data from fMRI, where features grow exponentially as the number of parcels increase, we might observe a very different relationship (ML Pipelines will scale different according to number of features). These analyses further considered only data as projected onto a surface as opposed to working with volumetric data directly. It would be interesting to repeat the experiments conducted here, but in native volumetric space or with different modalities of fMRI.
Another built inbuilt-in assumption in the current study was the decision to simply concatenate regions of interest from each of the four sub-modalities (e.g., thickness, myelination, ect…) into one input feature array. Possible extensions could focus on studying the impact of each modality as treated independently. This could lead to some interesting comparisons in performance between the different sub-modalities. Likewise, an extension could instead focus on choice of strategy for combining information from the different sub-modalities. For example, would performance increase if sub-modality specific models were trained and then combined via ensembled? Or- would other multi-modal fusion strategies work better?
It may be interesting in future work to investigate how stable the observed parcellation performance scaling is under different sample sizes, as Neuroimaging based ML in particular has been shown to respect a separate performance scaling with respect to sample size (Schulz 2020). Notably in this work, the 5-fold cross validation simulated roughly training models with a sample size of roughly 7,500, a hardly realistic sample size for many studies. Other options may include testing the current work with a range of different ML estimators. For example, deep learning based models may exhibit different scaling considering they can in theory better handle data with structured high dimensional feature spaces, though current work is conflicted on the merit of deep learning for neuroimaging based ML (He 2020, Abrol 2021).
Abrol, A., Fu, Z., Salman, M., Silva, R., Du, Y., Plis, S., & Calhoun, V. (2021). Deep learning encodes robust discriminative neuroimaging representations to outperform standard machine learning. Nature communications, 12(1), 1-17.
Schulz, MA., Yeo, B.T.T., Vogelstein, J.T. et al. Different scaling of linear models and deep learning in UKBiobank brain images versus machine-learning datasets. Nat Commun 11, 4238 (2020). https://doi.org/10.1038/s41467-020-18037-z