Development of an artificial intelligence-based recognition model for serum indices
10.3760/cma.j.cn114452-20250910-00510
- VernacularTitle:基于人工智能的血清指数识别模型构建
- Author:
Shenling LIAO
1
;
He HE
;
Xia WANG
;
Yifan ZHAO
;
Zhi LIU
;
Jin XU
;
Mei ZHANG
Author Information
1. 四川大学华西医院输血科，成都610041
- Publication Type:Journal Article
- Keywords:
Artificial intelligence;
Serum indices;
Lipemia;
Hemolysis;
Icterus
- From:
Chinese Journal of Laboratory Medicine
2025;48(12):1546-1551
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop an artificial intelligence-based model for automated recognition of serum indices using machine vision and deep learning.Methods:This study was a cross-sectional study.Serum sample images were collected fromWest China Hospital of Sichuan University from September 21, 2020 to January 20, 2023 using the imaging device of the fully automated sample pre-processing system. A computer random number generator was used to randomly select one whole hour each day, and all serum sample images processed within that hour were included. After excluding samples with unqualified images and missing serum index results, a total of 5, 534 samples were included. These were divided into a training set and a test set in an 8∶2 ratio using Python random shuffle function, and 4, 458 samples were in the training set and 1, 076 samples were in the test set. After manual inspection, the serum regions were annotated using the MATLAB Image Labeler tool and converted into YOLO format, and a YOLO v5-based model was constructed for automatic serum region extraction. The actual values of lipemia index (L-index), hemolysis index (H-index), and icterus index (I-index) were measured using the automatic biochemical analyzerwith matched reagent kits. A serum index regression model was constructed based on the MobileNet v2 network using the PyTorch 1.10.0 framework. The grading performance of the model was evaluated using accuracy, Kappa coefficient, sensitivity, specificity, positive predictive value, and negative predictive value. Regression performance was assessed using root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R2), and Bland-Altman analysis.Results:The overall accuracy rates for grading L-index, H-index, and I-index were 98.88%, 95.26%, and 92.47%, respectively, with Kappa coefficients of 0.72, 0.72, and 0.59. For L-index, MAE was 5.11, RMSE was 9.77, and R2 was 0.78. For H-index, MAE was 5.18, RMSE was 8.99, and R2 was 0.89. For I-index, MAE was 1.13, RMSE was 3.01, and R2 was 0.71. Bland-Altman analysis showed that 95.5%, 95.1%, and 95.7% of the data points fell within the consistency intervals for L-index, H-index, and I-index, respectively.Conclusion:The study developed an artificial intelligence-based serum index regression modelto estimate serum indices with high efficiency and accuracy. It shows great potential for reducing laboratory costs, improving clinical testing efficiency, and promoting intelligent development in laboratory medicine.
