Machine Learning Based Single-Frame Super-Resolution Processing for Lensless Blood Cell Counting

Area of Science:

• Biomedical Engineering
• Optical Imaging
• Machine Learning

Background:

• Lensless imaging systems with microfluidics and CMOS sensors offer miniaturized point-of-care testing (POCT) for blood cell counting.
• Limited resolution in lensless systems necessitates advanced processing for improved cell detection and recognition.

Purpose of the Study:

• To investigate and compare machine learning-based single-frame super-resolution (SR) techniques for enhancing resolution in lensless blood cell counting.
• To evaluate the effectiveness of Extreme Learning Machine based SR (ELMSR) and Convolutional Neural Network based SR (CNNSR) in improving cell resolution and counting accuracy.

Main Methods:

• Development and comparison of two single-frame SR algorithms: ELMSR and CNNSR.
• Implementation of these algorithms on lensless blood cell counting prototypes using commercial and custom CMOS image sensors.
• Quantitative evaluation of resolution enhancement and cell counting accuracy against a commercial flow cytometer.

Main Results:

• A 4× improvement in cell resolution was achieved using the proposed SR methods.
• CNNSR demonstrated a 9.5% higher resolution enhancement performance compared to ELMSR.
• Cell counting results from the SR-enhanced lensless system showed strong agreement with a commercial flow cytometer.

Conclusions:

• Both ELMSR and CNNSR effectively improve resolution in lensless blood cell counting systems.
• CNNSR offers superior performance for resolution enhancement, making it highly suitable for POCT applications.
• These SR techniques hold significant potential for advancing low-cost, high-performance lensless blood analysis devices.