Globally, retinal disorders impact thousands of individuals. Early diagnosis and treatment of these anomalies might halt their development and prevent many people from developing preventable blindness. Iris spot segmentation is critical due to acquiring iris cellular images that suffer from the off-angle iris, noise, and specular reflection. Most currently used iris segmentation techniques are based on edge data and non-cellular images. The size of pigment patches on the surface of iris increases with eye syndrome. In addition, iris images taken in uncooperative settings frequently have negative noise, making it difficult to segment them precisely. The traditional diagnosis processes are costly and time-consuming since they require highly qualified personnel and have strict environments. This paper presents an explainable deep learning model integrated with a multiclass support vector machine to analyze iris cellular images for early pigment spot segmentation and classification. Three benchmark datasets Mile, UPOL and Eyes SUB were used in the experiments to test the proposed methodology. The experimental results are compared on standard metrics, demonstrating that the proposed model outperformed the methods reported in the literature regarding classification errors. Additionally, it is observed that the proposed parameters are highly effective in locating micro pigment spots on the iris surfaces.

中文翻译：

---

通过可解释的深度学习和多类支持向量机，通过虹膜细胞图像分析进行早期色素斑分割和分类。

在全球范围内，视网膜疾病影响着成千上万的人。这些异常的早期诊断和治疗可能会阻止其发展，并防止许多人患上可预防的失明。由于采集的虹膜细胞图像会受到虹膜偏角、噪声和镜面反射的影响，因此虹膜点分割至关重要。目前使用的大多数虹膜分割技术都是基于边缘数据和非细胞图像。虹膜表面色素斑块的大小随着眼部综合症的增加而增大。此外，在不合作的环境中拍摄的虹膜图像经常具有负噪声，使得难以精确分割它们。传统的诊断过程需要高素质的人员和严格的环境，成本高且耗时。本文提出了一种与多类支持向量机集成的可解释深度学习模型，用于分析虹膜细胞图像以进行早期色素斑分割和分类。实验中使用了三个基准数据集 Mile、UPOL 和 Eyes SUB 来测试所提出的方法。将实验结果与标准指标进行比较，表明所提出的模型在分类错误方面优于文献中报告的方法。此外，据观察，所提出的参数在定位虹膜表面上的微色素点方面非常有效。