ECLNet: Center localization of eye structures based on Adaptive Gaussian Ellipse Heatmap

Area of Science:

• Ophthalmology
• Medical Imaging Analysis
• Computer Vision

Background:

• Accurate localization of biological structures in medical images is crucial for clinical diagnosis and treatment.
• Current heatmap-based keypoint estimation methods use fixed Gaussian kernels, limiting adaptability to morphological variations in target regions.
• Developing adaptive methods is essential for improving the precision of medical image analysis.

Purpose of the Study:

• To develop a deep learning network, ECLNet, for accurate localization of eye-related structures (optic disc, fovea center, eye center) in medical images.
• To propose an Adaptive Gaussian Ellipse Heatmap (AGEH) method that utilizes gradient features to dynamically adjust heatmap morphology.
• To evaluate the performance of ECLNet with AGEH against state-of-the-art methods on public datasets.

Main Methods:

• A deep convolutional neural network, ECLNet, was designed for keypoint localization.
• The Adaptive Gaussian Ellipse Heatmap (AGEH) method was proposed to generate heatmaps that adapt to target region morphology using gradient information.
• The method was validated on the IDRiD dataset for optic disc and fovea center localization and the CATARACT dataset for eye center localization.

Main Results:

• ECLNet achieved a mean Euclidean Distance of 17.995 pixels for optic disc and 39.446 pixels for fovea center on the IDRiD dataset.
• The method localized the eye center with a mean absolute Position Error of 0.186±0.027 mm on the CATARACT dataset.
• Comparative analysis demonstrated superior performance of ECLNet with AGEH over existing state-of-the-art methods.

Conclusions:

• ECLNet incorporating the AGEH method provides a robust and accurate solution for localizing critical eye structures in medical images.
• The adaptive nature of AGEH effectively addresses the limitations of fixed-kernel methods, enhancing localization accuracy.
• The proposed approach shows significant potential for improving computer-aided diagnosis and treatment planning in ophthalmology.