AI-Powered Serum Quality Assessment
A two-step deep learning system combining YOLOv5 and ResNet-50 to identify hemolysis, icterus and lipemia before compromised serum samples enter downstream clinical testing.
Why serum quality matters before a result is reported
Hemolysis, icterus and lipemia can interfere with biochemical measurements. Visual inspection is still vulnerable to subjectivity and variability, creating a need for fast, reproducible pre-analytical quality control.
Hemolysis
Ruptured erythrocytes release intracellular components into serum. Analytes such as AST, LDH and potassium may therefore appear artificially elevated.
Icterus
Bilirubin and related pigments have spectral properties capable of interfering with photometric and peroxidase-coupled biochemical measurements.
Lipemia
Excess serum lipids and chylomicrons can increase light scattering and absorbance, compromising multiple laboratory measurements.
Built on real-world clinical sample distributions
The model-development dataset contained 21,000 serum samples. Eighty percent were assigned to training and twenty percent to validation. A separate 21,000-sample dataset was subsequently used for independent evaluation.
Primary dataset
Observed serum abnormalities
513 abnormal / rejected samples
286 abnormal samples
754 abnormal / rejected samples
One model finds the serum.
The next model understands it.
Instead of classifying an entire laboratory image directly, the system first isolates the biologically relevant serum region. ResNet-50 then performs multi-level HIL classification.
Serum Image
Sample images are automatically captured during preprocessing in the clinical laboratory workflow.
CLINICAL INPUTROI Detection
The model identifies the visible serum area while excluding irrelevant information such as labels, caps and background regions.
YOLOv5Image Processing
Extracted serum images are normalized and resized to 224 × 224 pixels for classification.
224 × 224HIL Classification
Deep residual features are used to identify hemolysis, icterus, lipemia and their severity grades.
RESNET-50Clinical Decision
The resulting severity class can support actions such as acceptance, advisory comments, further review or sample rejection.
QC OUTPUTDeep learning localized serum more reliably
The researchers compared YOLOv5-based ROI localization against a traditional computer-vision workflow based on manually engineered image-processing features.
Engineered image processing
YOLOv5 serum ROI detector
Performance on previously unseen clinical data
The final model was evaluated on an independent dataset that was not used during model training or hyperparameter tuning. The figures below show binary detection performance.
Hemolysis
Icterus
Lipemia
Consistency where visual judgment varies
Nine trained laboratory professionals independently classified representative serum samples under standardized conditions. Their results were compared with the AI system.
Exact match requires complete agreement with the reference category. A fuzzy match allows a difference of no more than one severity level.
From “normal” to “reject”
Rather than producing only a positive or negative flag, the system was designed around severity grades that correspond to practical laboratory decisions.
Hemolysis Index · H
Five clinical categoriesIcterus Index · I
Three clinical categoriesLipemia Index · L
Five clinical categoriesDesigned for a high-throughput laboratory workflow
The model was evaluated in a clinical preprocessing workflow using existing laboratory equipment rather than a separate experimental imaging system.
AI inference at laboratory speed
The study reports mean inference time below 100 milliseconds per serum sample, supporting deployment in a high-throughput workflow.
cobas® c 701
The research prototype was integrated with preprocessing associated with the Roche cobas c 701 module. Roche describes the platform as a high-throughput clinical chemistry analyzer capable of determining hemolytic, icteric and lipemic serum indices.
Two established computer-vision ideas, applied to laboratory medicine
The study combines object localization and deep residual classification rather than asking a single network to interpret the entire raw laboratory image.
YOLOv5
Used in the study to localize the serum region of interest. YOLO-family models are designed for efficient visual object detection, making them suitable for rapid image localization.
Technology context: UltralyticsResNet-50
A deep residual convolutional network used here for HIL classification after ROI extraction. Residual connections were originally developed to make much deeper neural networks easier to optimize.
Technology context: He et al. · CVPRClinical QC
The AI output is not presented simply as an image label. It maps visual assessment to graded categories intended to support actionable pre-analytical quality-control decisions.
Application context: clinical laboratory workflowWhat the results establish — and what remains to be tested
Study limitations
The research was conducted as a single-center study.
Generalizability to other institutions, patient populations, instruments and pre-analytical workflows remains to be confirmed.
A prospective multicenter validation study is required before widespread clinical adoption.
Future work may compare the architecture with newer deep-learning models and explore continuous HIL index regression.
Why the research matters
The study moves serum-quality AI beyond a simple proof-of-concept. It evaluates a two-stage computer-vision workflow with clinically collected samples, an independent test dataset and direct comparison with trained laboratory professionals.
Its central contribution is the translation of visual serum inspection into a rapid, standardized and graded QC process that can potentially operate before compromised samples generate misleading downstream measurements.
From serum image to clinical quality signal.
Development and Evaluation of Artificial Intelligence-Based Two-Step Model for Automated Serum Quality Assessment in Clinical Laboratories.