Sclera Segmentation in Images for Bilirubin Level Measurement Using the U-Net Network

The sclera is a white membrane rich in collagen and elastin fibres, which gives it an affinity for bilirubin. This yellow substance is produced by the breakdown of the heme group, and when its levels are elevated, it causes jaundice a condition that leads to yellowing of the skin, mucous membranes, and sclera. The intensity of the yellowing in the sclera is directly related to bilirubin levels in the body. This relationship enables the extraction of features to infer these levels using machine learning techniques based on RGB images of the sclera. Sclera segmentation is a transcendent factor in achieving this goal. For this reason, this article presents the results of sclera segmentation using the U-Net network. This is a convolutional network composed of encoding and decoding layers and is used for medical image segmentation. The model was trained and validated with a set of 181 eye images and their corresponding binary masks. The results obtained during the training phase are Loss (0.006), Precision (0.976), Recall (0.973) and F1-score (0.974), and in the validation phase: Loss (0.145), Precision (0.897), Recall (0.863) and F1-Score (0.880). These results demonstrate the U-Net model's effectiveness in segmenting the sclera, particularly in the training phase where the metrics are highly favorable. However, the slight decrease in performance during the validation phase suggests the need for further refinement. Future work will focus on increasing the dataset size and introducing data augmentation techniques to improve generalization and robustness. Ultimately, accurate sclera segmentation is a critical step toward developing reliable Machine Learning models for non-invasive bilirubin level estimation.