A deep learning-based system for assessment of serum quality using sample images.

ABSTRACT

BACKGROUND: Serum quality is an important factor in the pre-analytical phase of laboratory analysis. Visual inspection of serum quality (including recognition of hemolysis, icterus, and lipemia) is widely used in clinical laboratories but is time-consuming, subjective, and prone to errors. METHODS: Deep learning models were trained using a dataset of 16,427 centrifuged blood images with known serum indices values (including hemolytic index, icteric index, and lipemic index) and their performance was evaluated by five-fold cross-validation. Models were developed for recognizing qualified, unqualified and image-interfered samples, predicting serum indices values, and finally composed into a deep learning-based system for the automatic assessment of serum quality. RESULTS: The area under the receiver operating characteristic curve (AUC) of the developed model for recognizing qualified, unqualified and image-interfered samples was 0.987, 0.983, and 0.999 respectively. As for subclassification of hemolysis, icterus, and lipemia, the AUCs were 0.989, 0.996, and 0.993. For serum indices and total bilirubin predictions, the Pearson's correlation coefficients (PCCs) of the developed model were 0.840, 0.963, 0.854, and 0.953 respectively. Moreover, 30.8% of serum indices tests were deemed unnecessary due to the preliminary application of the deep learning-based system. CONCLUSIONS: The deep learning-based system is suitable for the assessment of serum quality and holds the potential to be used as an accurate, efficient, and rarely interfered solution in clinical laboratories.