Histogram-based thresholding for detection and quantification of hemorrhages in retinal images

Abstract EN

Retinal image analysis is commonly used for the detection and quantification of retinal diabetic retinopathy.

In retinal images, dark lesions including hemorrhages and microaneurysms are the earliest warnings of vision loss.

In this paper, new algorithm for extraction and quantification of hemorrhages in fundus images is presented.

Hemorrhage candidates are extracted in a preliminary step as a coarse segmentation followed by a fine segmentation step.

Local variation processes are applied in the coarse segmentation step to determine boundaries of all candidates with distinct edges.

Fine segmentation processes are based on histogram thresholding to extract real hemorrhages from the segmented candidates locally.

The proposed method was trained and tested using an image dataset of 153 manually labeled retinal images.

At the pixel level, the proposed method could identify abnormal retinal images with 90.7% sensitivity and 85.1% predictive value.

Due to its distinctive performance measurements, this technique demonstrates that it could be used for a computer-aided mass screening of retinal diseases