The orininal code for project "AMD Neovascular activity prediction using OCT- angiography based on entropy and deep learning" under the College Student Research Scholarship(National Science Council, Taiwan)
The results now assists ophthalmologists at TPEVGH in interpreting OCTA images (patent pending), significantly advancing the field.
the OCTA_Net model is modified from https://github.com/Luodian/Otter](https://github.com/iMED-Lab/OCTA-Net-OCTA-Vessel-Segmentation-Network.git)https://github.com/iMED-Lab/OCTA-Net-OCTA-Vessel-Segmentation-Network.git
The study aims to explore the potential of incorporating the information science concept of entropy into the classification of neovascular exudative and quiescent eyes in age-related macular degeneration (AMD).
A total of 165 optical coherence tomography angiography (OCTA) image pairs (69 reactive events and 96 treatment events) were collected from an AMD follow-up cohort at Taipei Veterans General Hospital. After following previous literature to employ Gabor filters for image augmentation, we processed the superficial capillary plexus (SCP) into OCTA vascular density maps, centerline maps, and foveal avascular zone (FAZ) masks. Correlation analysis was performed on OCTA metrics, including entropy, vessel density, vessel caliber, vessel tortuosity, fractal dimension, FAZ area, and FAZ circularity. Additionally, a supervised machine learning algorithm, the eXtreme Gradient Boost (XGBoost) classifier, was developed to categorize images into exudative and quiescent AMD groups.
Our two-tailed paired t-test revealed that the entropy, vessel density, and fractal dimension increased significantly (p < 0.05) in both reactive and treatment events, while vessel tortuosity showed significant increases only during treatment events. After applying the Gabor filter, additional increases in vessel caliber (reactive) and vessel tortuosity (treatment) were observed (p < 0.05). However, the FAZ area and circularity did not reach statistical significance for either event type. Compared to a baseline classifier with an accuracy of 0.867 and AUROC of 0.837 (sensitivity of 0.95, and specificity of 0.72), our entropy-integrated XGBoost classifier demonstrated a performance improvement of over 10%, achieving an accuracy of 0.967 and AUROC of 0.967 (sensitivity of 0.93, and specificity of 1.00).
Our study indicates that incorporating entropy into the evaluation of OCTA metrics enhances the classification of exudative and quiescent AMD. This improvement could contribute to more accurate diagnoses and better management of the condition.
