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

Visual assessment of serum quality in clinical labs is common but has limitations. It requires manual inspection for conditions like hemolysis, icterus, and lipemia. This process is subjective and time-consuming, leading to variability in outcomes. Prior research has shown that deep learning can automate similar diagnostic tasks. However, no prior work had resolved how to apply deep learning to serum quality assessment using image data. This gap motivated the development of a system that could reduce human error and increase efficiency. The need for an objective, rapid, and accurate method became clear. No prior work had resolved how to integrate multiple serum indices into a single automated system. The challenge was to create a model that could distinguish between qualified and unqualified samples. This uncertainty drove the exploration of deep learning models for this purpose.

The aim of this study was to develop a deep learning-based system for automatically assessing serum quality using sample images. The specific problem addressed was the inefficiency and subjectivity of manual visual inspection in clinical labs. The motivation stemmed from the need to reduce errors and time spent on serum quality assessment. The system needed to recognize hemolysis, icterus, and lipemia accurately. The goal was to integrate multiple serum indices into one automated solution. The researchers proposed using a dataset of centrifuged blood images with known serum indices. The system was designed to classify samples into qualified, unqualified, or image-interfered categories. The study aimed to evaluate the system's performance using cross-validation.

The study used a dataset of 16,427 centrifuged blood images with known serum indices. Deep learning models were trained using these images to recognize hemolysis, icterus, and lipemia. Five-fold cross-validation was used to evaluate model performance. The models were developed to classify samples into three categories: qualified, unqualified, and image-interfered. Predictive models were also trained to estimate serum indices and total bilirubin levels. The system combined these models into a unified deep learning-based solution. The models were tested for accuracy using area under the ROC curve (AUC) and Pearson's correlation coefficients (PCC). The study focused on the integration of multiple diagnostic tasks into a single system.

The deep learning model achieved an AUC of 0.987 for identifying qualified samples. For unqualified samples, the AUC was 0.983. The model showed an AUC of 0.999 for image-interfered samples. Hemolysis subclassification had an AUC of 0.989. Icterus subclassification reached an AUC of 0.996. Lipemia subclassification had an AUC of 0.993. The model predicted serum indices with PCCs of 0.840, 0.963, 0.854, and 0.953. The system reduced unnecessary serum indices tests by 30.8%. These results suggest the model can accurately assess serum quality. The system demonstrated high performance across multiple diagnostic tasks.

The study concludes that the deep learning-based system can effectively assess serum quality. The system accurately classifies samples into qualified, unqualified, and image-interfered categories. The authors propose that the system can reduce manual inspection time and errors. The model's high AUC values suggest strong diagnostic performance. The system's ability to predict serum indices with high PCCs was noted. The reduction of 30.8% in unnecessary tests was highlighted as a benefit. The authors suggest the system is suitable for clinical use. The study emphasizes the potential of deep learning in automating laboratory diagnostics.