Improved classification and grading of interferents in serum specimens using machine learning

Serum specimens containing interferents affect the accuracy of test results through a variety of mechanisms. Consequently, rigorous quality control of serum samples before biochemical analysis may help prevent incorrect results. Based on the hypothesis that serum color images contain information about the category and concentration of interferents, a machine learning method was proposed to automatically classify and grade color images of serum samples into three categories and five levels of interferent concentration. First, using a color correction method, the color image was preprocessed to eliminate ambient light color cast during the shooting process. Serum regions were then segmented using a convolutional neural network. Subsequently, color moment features were extracted and utilized in the classification of hemolysis, icterus, and lipemia (HIL), the three most common interferents in blood examinations. Finally, feature selection was utilized to select the most suitable features for grading the degree of HIL. This feature subset was used to grade five concentration levels for each category. The Dice coefficient and IoU of the serum region segmentation results were 96.36% and 93.02%, respectively. The accuracy and F1-score for classification were both 1. For the grading task, the accuracies were 0.9829, 0.9876, and 0.9526, and Fl-scores were 0.9828, 0.9876, and 0.9520 for hemolysis, icterus, and lipemia, respectively. The proposed method can successfully identify if a sample contains HIL interference and grade the degree of interferent concentration, providing an efficient and feasible method for serum quality control.