Wyniki wyszukiwania - BazTech

Ograniczanie wyników

1 Biocybernetics and Biomedical Engineering

1 2019

Znaleziono wyników: 1

Liczba wyników na stronie

Wyniki wyszukiwania

Wyszukiwano:
w słowach kluczowych: unsupervised feature learning

Sortuj według:

Ogranicz wyniki do:

Combination of hand-crafted and unsupervised learned features for ischemic stroke lesion detection from Magnetic Resonance Images

Bharathi Praveen Gurunath, Agrawal Anita, Sundaram Ponraj, Sardesai Sanjay

Biocybernetics and Biomedical Engineering

2019

Vol. 39, no. 2

410--425

Detection of ischemic stroke lesions plays a vital role in the assessment of stroke treatments such as thrombolytic therapy and embolectomy. Manual detection and quantification of stroke lesions is a time-consuming and cumbersome process. In this paper, we present a novel automatic method to detect acute ischemic stroke lesions from Magnetic Resonance Image (MRI) volumes using textural and unsupervised learned features. The proposed method proficiently exploits the 3D contextual evidence using a patch-based approach, which extracts patches randomly from the input MR volumes. Textural feature extraction (TFE) using Gray Level Co-occurrence Matrix (GLCM) and unsupervised feature learning (UFL) based on k-means clustering approaches are employed independently to extract features from the input patches. These features obtained from the two feature extractors are then given as input to the Random Forest (RF) classifier to discriminate between normal and lesion classes. A hybrid approach based on the combination of TFE using GLCM and UFL based on the k-means clustering is proposed in this work. Hybrid combination approach results in more discriminative feature set compared with the traditional approaches. The proposed method has been evaluated on the Ischemic Stroke Lesion Segmentation (ISLES) 2015 training dataset. The proposed method achieved an overall dice coefficient (DC) of 0.886, precision of 0.979, recall of 0.831 and accuracy of 0.8201. Quantitative measures show that the proposed approach is 28.4%, 27.14%, and 5.19% higher than the existing methods in terms of DC, precision, and recall, respectively.