Severity estimation of brainstem in dementia MR images using moth flame optimized segmentation and fused deep feature selection

Dementia is an irreversible chronic neuro-disorder. Prediction of preclinical stages variation in dementia disorder helps to delay the progression. This study attempts to differentiate the brainstem (BS) structures for normal and different demented stages. BS structure is interconnected with many brain cortical structures that help to provide valuable pathological information about atrophy. This work is carried out using ADNI public database. Initially, skull-stripped MR images are used to perform BS segmentation using moth flame optimization-based multilevel Tsallis entropy method. Architecture such as AlexNet, GoogleNet and SqueezeNet is considered to extract features. The fusion of these features provides optimal information about BS structures for preclinical stages. The performance of fused deep features is evaluated using heat map and occlusion sensitivity map. Further, combined feature selection is carried out to extract the most distinct feature set using mutual information, minimum redundancy maximal relevance and recursive feature elimination methods. Finally, the analysis of variance is used to evaluate inter and intra-class variation of the subject. Results indicate that the suggested approach could segment the BS prominently. The correlation value was found to be > 0.97 in all the considered stages. The heatmap and occlusion sensitivity show the fused deep features provide highly discriminative features. The statistical performance of this fused feature set of Normal (CN)/EMCI, CN/EMCI, CN/LMCI, CN/MCI, CN/AD, EMCI/MCI, EMCI/AD, MCI/LMCI, MCI/AD and LMCI/AD shows high significant variation (p < 0.0001). Consequently, this approach captures the complex preclinical stage variation effectively and suitable to reduce the misdiagnosis rates.