Transfer-transfer model with MSNet: An automated accurate multiple sclerosis and myelitis detection system

Transfer-transfer model with MSNet: An automated accurate multiple sclerosis and myelitis detection system

Purpose
Multiple sclerosis (MS) is a commonly seen neurodegenerative disorder, and early diagnosis of MS is a crucial issue to promote patient health. Since MS diagnosis is a computer vision problem, machine learning can be utilized for this purpose. Important research has used transfer learning (TL) to rapidly apply the advantages of deep learning models. Therefore, TL has developed a wide usage in computer vision applications. Herein, we describe a new algorithm in this regard, termed transfer-transfer (TT). To implement the algorithm, a multi-result machine learning model is required. In order to determine efficacy, we use transfer learning-based and hybrid feature engineering. The goal is to demonstrate the classifiability of the TT model.

Materials and method
A new magnetic resonance image dataset containing three classes were collected to obtain TT model results, i.e.: (1) multiple sclerosis (MS), (2) myelitis, and (3) control patients. We have designed this model for MS detection. Thus, we named it MSNet. For deep feature engineering, MSNet with two layers of pretrained DenseNet201 and two layers of ResNet50 was incorporated into the system since these networks are highly accurate. By deploying these four layers, four feature vectors were calculated. ReliefF (RF), Chi2, and Neighborhood Component Analysis (NCA) were utilized in the feature selection phase, and the number of the feature vectors is increased from 4 to 12 (=4 × 3). By using k-nearest neighbor (kNN) and support vector machine (SVM) classifiers, 24 (=12 × 2) outputs were calculated, with the best result created by applying information fusion. TT incorporates the information fusion findings to construct a new feature vector. The most salient features were selected by deploying an iterative feature selector, and the features chosen were then classified.

Results
The TT-based MSNet was applied to the magnetic resonance (MR) image dataset, yielding a 97.63% classification accuracy.

Conclusions
The findings and computed results demonstrate that the TT model with MSNet outperforms existing systems with increased transfer learning classifiability.