研究目的
To develop a computer-aided diagnostic method for stroke prediction using retinal fundus images by implementing Histogram of Oriented Gradients (HoG) for feature extraction and comparing it with Haralick features, with the aim of achieving high accuracy in classifying normal and stroke-affected images.
研究成果
The HoG-based feature extraction method achieved a high accuracy of 93% and an AUC of 0.979 in classifying retinal images for stroke prediction, outperforming Haralick features (87% accuracy). This non-invasive approach can assist physicians in early stroke diagnosis by analyzing retinal vasculature changes. Future work should involve larger datasets and additional features to improve predictive accuracy.
研究不足
The dataset is limited to 130 images from a single source, which may affect generalizability due to data privacy constraints. The method relies on manual localization of the optic disc center, which could introduce human error. Comparison with existing state-of-the-art methods is not possible as private datasets were used in prior works. The approach is computationally less intensive than deep learning but may not capture all complexities without a larger dataset.
1：Experimental Design and Method Selection:
The study uses a comparative approach between HoG and Haralick features for texture analysis in retinal images to predict stroke. HoG is chosen for its ability to capture fine details and invariance to illumination changes.
2：Sample Selection and Data Sources:
Retinal fundus images were collected from Sree Gokulam Medical College and Research Foundation, Kerala, India, with 80 normal and 50 stroke patient images (total 130 images), each of size 2336x3504 pixels. Ethical approval was obtained (SGMC-IEC No. 25/293/01/2017).
3：7). List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: A fundus camera was used for image acquisition. Software for image processing and feature extraction (specific models or brands not mentioned).
4：Experimental Procedures and Operational Workflow:
Images were converted to grayscale; the optic disc center was manually located; a region of interest (ROI) of size 256x256 pixels was cropped around the optic disc center (Zone C). HoG features were extracted using 8x8 cells, 2x2 cells per block, and 9 orientation bins, resulting in a feature vector of size 1x8100 per image. Haralick features (14 types) were also computed. Features were input to a Na?ve Bayes classifier, with 10-fold cross-validation for evaluation.
5：Data Analysis Methods:
Performance metrics included accuracy, Kappa statistic, root mean square error (RMSE), and area under the ROC curve (AUC). Statistical analysis (e.g., regression for feature selection) was performed to identify significant Haralick features.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
