A deep learning approach to gold nanoparticle quantification in computed tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nanotechnology

Background:

• Deep learning (DL) is increasingly applied to medical imaging analysis.
• Gold nanoparticles (AuNPs) have potential biomedical applications requiring precise detection and quantification.
• Clinical multi-detector CT (MDCT) scanners present challenges for imaging low and high concentrations of nanoparticles.

Purpose of the Study:

• To evaluate the efficacy of DL for classifying and quantifying gold nanoparticles (AuNPs) using a clinical MDCT scanner.
• To determine if DL can detect AuNPs at concentrations below human visual perception.
• To assess DL's capability in quantifying high AuNP concentrations that exceed the dynamic range of conventional CT.

Main Methods:

• AuNPs were imaged in a human-sized phantom across a wide concentration range (0.0274–200 mgAu/mL).
• A convolutional neural network (CNN) was trained and tested on CT image slices acquired with standard clinical parameters (120 kVp, 23.6 mGy CTDIvol).
• The CNN was designed to classify 17 different tissue types and varying AuNP concentrations.

Main Results:

• DL achieved 95% accuracy in classifying AuNPs at 0.1095 mgAu/mL and 97% accuracy at 0.2189 mgAu/mL, both below human visual detection limits.
• Accurate classification (95%) was also achieved for high AuNP concentrations (150 and 200 mgAu/mL).
• High concentrations produced CT numbers at or exceeding the 12-bit dynamic range limit, necessitating specialized analysis.

Conclusions:

• DL effectively detects gold nanoparticles (AuNPs) at concentrations imperceptible to the human eye.
• The study demonstrates DL's ability to quantify very high AuNP concentrations that surpass the dynamic range of clinical MDCT scanners.
• This DL-based quantification method may extend to other high-density objects in CT imaging, potentially eliminating the need for extended Hounsfield scales.