Automated ischemic acute infarction detection using pre-trained CNN models' deep features. (April 2023)

• Record Type: Journal Article
• Authors: Tasci, Burak
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: Pre-trained models -- SVM -- IMV -- Ischemic acute infarction detection; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104603
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Acute ischemic infarcts dataset(v2) was collected. In the proposed method, a total of 22 layers of the Efficientb0, DenseNet201, Resnet101, Resnet50, Inceptionresnetv2, Xception, MobileVnet2, ShuffleNet, Darknet19, NasnetLarge, and AlexNet pre-trained models were used. High accuracy was achieved with ImRMR and IMV. Our model is tested on a different four brain image dataset. Over 97.90 % accuracy was obtained with the proposed method in all datasets. Abstract: Background: Cerebrovascular Diseases (CVD) constitute more than 50 % of neurological diseases requiring hospital treatment. Stroke is a type of disease that causes physical disability to survivors; approximately 85 % of strokes are ischemic. Once the symptoms have been identified, the diagnosis of the disease can be confirmed using widely available imaging techniques. This study attempts to build a model using diffusion Magnetic Resonance Imaging (MRI) to diagnose stroke. Materials and method: Four different datasets were used in the study. A total of 1112 acute ischemic infarctions and 1202 normal diffusion MR images were collected for this study. In addition, two datasets had two classes, and the other dataset had four classes. These MRI datasets were used to test and compare the method. Efficient b0, DenseNet201, Resnet101, Resnet50, Inceptionresnetv2, Xception, MobileVnet2, ShuffleNet, Darknet19, NasnetLarge, AlexNet pre-trained Convolutional Neural Network (CNN) models with over 90 % classification results were used in the study. Features were obtained from the softmax/classification layers of these CNN models. The data were then classified utilizing a Support Vector Machine (SVM) classifier with 10-fold Cross-Validation (CV). The prediction vectors obtained from the classifiers were sorted and applied to the Iterative Majority Voting (IMV). Results: The proposed method yielded 97.93 %, 97.98 %, 99.07 %, and 96.93 % results for accuracy, F1-Score, precision, and recall for the collected dataset, respectively. For Dataset 2, 99.32 % accuracy and 100 % F1-Score, precision, and recall 100 % results were obtained. For Dataset 3, 97.96 % accuracy and 98.37 % F1-Score, 99.42 % precision, and 99.02 % recall results were obtained. For Dataset 4, accuracy, F1-Score, precision, and recall 100 % results were obtained. Conclusions: In line with the results obtained with the method proposed in this study, acute ischemic infarction can be successfully classified based on diffusion MRIs. This research succeeded because it achieved high-accuracy results in four different datasets. This advanced automated system can assist neurologists in determining whether they miss anything when detecting acute ischemic infarction, thereby minimizing potential human error.