The Hybrid Transfer Learning Framework incorporating DaTSCAN SPECT Imaging to Enable Differential Diagnosis of Parkinson’s Disease with the SWEDD Group

Abstract

With the ability to enhance patient treatment and diagnosis, early Parkinson disease (PD) diagnosis is currently one of the top priorities in the medical sector.  Early medical action and disease treatment are facilitated by prompt and precise diagnosis.  A group of Scan Without Evidence of Dopaminergic Deficit (SWEDD) refers to individuals with mixed clinical features and imaging results from the two cohorts of people, Parkinson's disease (PD) and healthy controls.  It might be challenging to detect Parkinson's disease (PD) in these hybrid instances, which further calls for accurate diagnosis and the application of image analysis. The current study has explicitly explored various deep learning and machine learning techniques to increase diagnosis accuracy in an effort to overcome the aforementioned issues. An Ensemble transfer learning models were specifically evaluated to reliably differentiate between individuals with PD, healthy controls, and SWEDD individuals. The data was obtained from PPMI, and we use DaTSCAN single-photon emission computed tomography (SPECT) scans for 457 subjects, which we classify as: 171 PD, 150 healthy controls, and 136 SWEDD individuals. Due to the limited number of images to construct our database system, we incorporated a Simple Generative Adversarial Network (GAN) image generation methods to introduce additional new subject images, leading to a total of 300 new images to be ( 100 for each category) for all the 3 categories of PD, HC and SWEDD incorporated. GAN augmentation was applied only to the training set of images.The proposed DNN model was then applied on this combined dataset of original PPMI and GAN generated images. Hybrid transfer learning models like DenseNet121+SVM, ResNet50 + SVM, ResNet152+ SVM, Xception +SVM, ResNet101+SVM etc were applied on balanced dataset in order to establish robustness and the ability of the model to be generalized. We have considered classification metrics like accuracy, recall, precision and F1_score performance comprehensively for assessing the performance of each model. We found that the proposed DNN model + Random Forest performed better with distinguishing scores of 80% accuracy whereas, the hybrid transfer learning models like DenseNet121+SVM, ResNet50 + SVM, ResNet152+ SVM, Xception +SVM gives 79%, 83.5%, 77%, 84% accuracy respectively. Among all the models, Xception +SVM gave better performance with 0.83 precision, 0.83 recall, 0.83 F1_score and 84% Accuracy. These are average of all values for 3 categories PD, HC and SWEDD, and were higher than the estimation in the conventional ML/DL algorithms. The Xception +SVM model has returned better results also for Class 0-HC as 0.80 F1_score, for Class1-PD 0.94 F1_score and for Class 2-SWEDD, as 0.75 F1_Score.  These results indicate the reliability of the proposed configuration of deep learning to establish the detection of PD cases with healthy ones and the population of SWEDD individuals, which is a milestone in early-stage PD diagnosis