A hybrid feature extraction approach for brain MRI classification based on Bag-of-words. (February 2019)

• Record Type: Journal Article
• Authors: Ayadi, Wadhah; Elhamzi, Wajdi; Charfi, Imen; Atri, Mohamed
• Is Part Of: Biomedical signal processing and control. Volume 48(2019)
• Page Start: 144
• Page End: 152
• Publication Date: 2019-02
• Subjects: BoW -- Brain tumor -- CAD -- Classification -- DWT -- MRI; 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.2018.10.010
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: This paper discusses different techniques for Magnetic resonance imaging (MRI) classification. Two techniques have been used: Discrete Wavelet Transform (DWT) and Bag-of-Words (BoW). A comparative study with several works showed efficiency and robustness of our system. Abstract: Magnetic resonance imaging (MRI) has attracted considerable attention in medical engineering community, since it is a non-invasive diagnostic technique and for its importance in medicine applications. With the aim to study and interpret the image more clearly, a computer-aided diagnosis (CAD) is required. Many automatic classification methods are proposed to classify the human brain MRI (normal/abnormal) to enhance the classification time and decrease the human error. This paper discusses different techniques for MR image classification where different tools are used for features extraction and classification. Based on these reviewed techniques, a new scheme is proposed. Our technique system exploits the benefits of two techniques: Discrete Wavelet Transform (DWT) and Bag-of-Words (BoW). For the validation step, we carried out several experiments based on 256 × 256 images from three datasets (DS-66, DS-160, DS-255) provided by Harvard Medical School. We applied 10 repetitions of k-fold stratified Cross Validation (CV) technique to validate the system performance. The Accuracies reached are respectively 100%, 100%, and 99.61% for DS-66, DS-160, and DS-255 datasets. The overall computation time is about 0.027 s for each MR image. A comparative study with several works showed efficiency and robustness of our system.