Towards machine learned quality control: A benchmark for sharpness quantification in digital pathology.

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Imaging Quality Control

Background:

• Pathology is transitioning to a digital, quantitative discipline driven by high-throughput scanning.
• Quality control (QC) of digital slides is a bottleneck, with manual screening for blur being time-consuming.
• Existing scanner software limitations necessitate robust blur detection methods.

Purpose of the Study:

• To develop and evaluate an automated blur detection framework for digital pathology slides.
• To compare machine learning approaches (feature-based and deep learning) against human experts for sharpness assessment.
• To provide a pipeline for quantifying and localizing blurred regions on digital slides.

Main Methods:

• Creation of a benchmark dataset for blur detection in digital pathology.
• Comparison of state-of-the-art sharpness descriptors and random forest models.
• Training of convolutional neural networks (residual networks) for blur detection.
• Evaluation of feature-based and deep learning methods for sharpness classification and regression.
• Human perception study to compare automated methods with domain experts.

Main Results:

• Achieved high accuracy in sharpness classification (99.74%) and regression (MSE 0.004) using machine learning models.
• Developed a pipeline generating spatial heatmaps to quantify and localize blurred areas.
• Demonstrated superior performance over current QC pipelines, reducing the error rate from 17% to 4.7% in a clinical setting.

Conclusions:

• The proposed automated blur detection framework significantly improves QC efficiency and accuracy in digital pathology.
• Machine learning, particularly deep learning, offers a robust and scalable solution for assessing digital slide quality.
• This framework has the potential to enhance the reliability and throughput of digital pathology workflows.