Detection and classification of progressive supranuclear palsy from MRI images using deep learning

Abstract

A timely and reliable computer aided MR image-based evaluations is required for the detection and classification of Progressive supranuclear palsy (PSP). It is a neurodegenerative ailment that is clinically very difficult to identify due to a high degree of overlap in characterized symptoms. Unlike other ailments, the primary constraint regarding PSP is the limited research work in this area. The main aim of our study is to establish a generalized model by comparing traditional custom CNN and transfer learning models such as DenseNet121, ResNet50, InceptionV3, VGG16, EfficientNetB0, Xception, MobileNet and InceptionResNetV2 which are fine tuned for the detection and classification of PSP with higher accuracy rates. Existing research has primarily focused on reducing the time complexity of neural networks and has only had success with low-level features. Furthermore, obtaining a significant volume of distributed labelled data is difficult. In our research, 125 T1 protocol based high resolution MRI images of 65 PSP and 60 normal control patients were considered. The image dataset is pre-processed, normalized and augmented before deploying them to the respective networks. The results propose that ‘InceptionResNetV2’ model can be considered as a generalised model for the detection and classification of PSP. The network offered a classification accuracy of 95%, Precision of 100%, Sensitivity of 92.8%, F1 score of 96.2% and a Specificity of 100% which is significantly higher compared to other models considered in the study and also the existing conventional ML models, thereby providing a prerequisite for significant diagnostic implementation.