MRomicsNet: A morphomics-radiomics-driven adaptive topological model for AD diagnosis on clinically routine...

Both morphomics and radiomics are typical features when constructing brain networks on the clinically routine T1-weighted images (T1WI) in Alzheimer's disease (AD) diagnosis but the integration was rarely reported. The study was to devise a morphomics-radiomics-driven adaptive topological model (MRomicsNet) by leveraging the strengths of morphomics and radiomics in AD representation.

An experimental validation between morphomics and radiomics was conducted to clarify the individual strengths in capturing brain topology and brain region features. The MRomicsNet was then conducted by integrating the individual strengths via a deep learning framework. Specifically, the MRomicsNet consisted of a morphomics-based graph convolution channel (morphGCN channel) and a morphomics-radiomics-based graph convolution channel (mrGCN channel). The morphGCN channel was fed with morphomics-based brain network to establish a sparse brain topology by strengthening important inter-regional connections while suppressing irrelevant ones. Given the generated sparse topology and the brain region features by radiomics, the new graph was then established and fed into the mrGCN channel for AD diagnosis.

The MRomicsNet was validated on the ADNI and EDSD datasets with a 10-fold cross-validation strategy. The experimental results demonstrated that the MRomicsNet achieved a relatively large improvement of 5.27% to 13.79% on the ADNI dataset and 3.70% to 10.22% on the EDSD dataset in diagnostic accuracy when compared to the existing methods. The complementary mechanism of the morphomics- vs radiomics-based brain network in AD representation was also validated by a visualization result.

The MRomicsNet highlighted the potential value of the morphomics and radiomics integration in facilitating T1WI-based structural brain network establishment and its associated AD diagnosis.