An intensity-aware vision transformer framework for precise localization of vitreous hemorrhage in fundus imaging

Vitreous hemorrhage is a serious ocular condition that requires accurate and timely diagnosis. Traditional methods for identifying affected areas in fundus images have limitations in precision and reliability. There is a need for advanced techniques to enhance the detection of vitreous hemorrhage and improve patient outcomes.

Detecting hemorrhages in color fundus images is difficult due to elements such as noise, feature intensity, and overlapping areas. In particular, the variation in feature intensity between different sections of the retina and lens leads to uniform (homogeneous) characteristics.

This study introduces a Feature Intensity Vision Transformer (FIVT) model utilizes vision transformer technology to detect high-intensity regions within segmented patches based on minimal pixel distributions. The model follows a systematic pipeline involving input image pre-processing, patch extraction (1 × 1-32 × 32), feature intensity estimation using the gray-level co-occurrence matrix (GLCM), and Vision Transformer-based intensity classification to produce a precise region detection mask. Feature distribution is analyzed to identify maximum intensity areas. High and low-intensity regions are classified, and pixels with significant feature content are matched with trained inputs. The model is trained using external images and high-feature regions until accuracy stabilizes.

We utilize the "exudate-hemorrhage-health" retinal image dataset to evaluate the proposed FIVT. This dataset comprises over 3,000 fundus color images, which include both internal and external references. Out of these, 536 images showing infections are employed to examine the infected areas. For the training process, 3,000 images are categorized into healthy, hemorrhage, and exudate groups.

Compared to other methods, the FIVT model achieved a 12.13% improvement in detection accuracy, a 12.55% increase in specificity, and a 12.32% enhancement in sensitivity, demonstrating robust performance across varying illumination and feature intensity conditions.

The model effectively handles varying illumination and intensity distributions, improving the precision of vitreous hemorrhage localization. However, its performance can be further enhanced with larger, multi-institutional datasets and integration of 3D volumetric data. Future work will focus on clinical validation and adaptation of the framework for related retinal disorders such as diabetic retinopathy and macular edema.