An improved YOLOv5 algorithm for automatic scleral plaque and pigmentation detection

The biological characteristics of the human eye are valuable in painless and non-invasive disease diagnosis, as they can be readily observed and accessed externally. Among these features, the scleral plaques and pigmentation, which are directly visible on the ocular surface, can provide useful information on an individual's health status. However, due to the inherent variability in the size and the shape of plaques and pigmentation spots among patients, automatic detection of these features in scleral images remains a significant challenge that has yet to be fully explored. In this study, we develop an object detection algorithm based on YOLOv5 for automatic detection of plaques and pigmentation spots in the scleral image. Specifically, the scleral region is initially extracted by the UNet++ network, and then an improved YOLOv5 model equipped with a convolution block attention module and Mamba blocks is employed to detect potential plaques and pigmentation spots in the scleral region. The convolution block attention module facilitates the detection and characterization of targets by eliminating redundant information, whereas the Mamba block integrates the state space model to expand the receptive fields of the proposed model, thereby enhancing its performance. The proposed detection algorithm was validated on a clinical dataset comprising scleral images from 388 subjects. The experimental results demonstrate that the proposed algorithm is capable of effective detection of scleral plaques and pigmentation spots and that it outperforms the original YOLOv5 model as well as other competing object detection algorithms.