Leukocyte super-resolution via geometry prior and structural consistency.

Area of Science:

• Medical Imaging
• Computational Biology
• Artificial Intelligence

Background:

• Single-image super-resolution (SISR) using deep convolutional neural networks has advanced significantly.
• Existing SISR methods struggle with leukocyte imaging due to insufficient exploration of leukocyte geometry and structural consistency.
• Inaccurate super-resolution (SR) results in leukocyte imaging can impede pathological studies by affecting cell type identification and subsequent analysis.

Purpose of the Study:

• To develop a deep network that effectively utilizes geometry prior and structural consistency for leukocyte image super-resolution.
• To establish and annotate a comprehensive leukocyte dataset for training models on structure and geometry.
• To improve the accuracy and reliability of SR in leukocyte imaging for enhanced pathological examination.

Main Methods:

• A novel deep network architecture comprising a prior network and an SR network was proposed.
• The prior network estimates a parsing map from the low-resolution (LR) image.
• The SR network integrates the estimated parsing map and the LR image to generate the high-resolution (HR) image.

Main Results:

• The proposed method significantly improves SR performance for leukocyte images, as evidenced by a 0.4 dB increase in peak-signal-to-noise ratio (PSNR) in benchmark tests.
• Experiments demonstrate that incorporating geometry prior and structural consistency enhances SR outcomes.
• The method outperforms state-of-the-art SR techniques in terms of PSNR and structural similarity indices.

Conclusions:

• The developed method accurately preserves the structural details of leukocytes, surpassing existing SR approaches.
• The proposed approach offers potential for wide-field cell prescreening applications using low-power objectives.
• This work advances the field of medical image analysis by providing a more robust solution for leukocyte super-resolution.