Using optimal transport theory to optimize a deep convolutional neural network microscopic cell counting method.

Area of Science:

• Medical image processing
• Computational biology
• Microscopy

Background:

• Accurate cell counting in microscopic images is crucial but challenging due to variations in cell size and shape.
• Existing methods often use deep learning (CNNs) or regression for cell counting, with limitations in precision.

Purpose of the Study:

• To develop a novel and more accurate cell counting method for microscopic images.
• To enhance the reliability of cell counting in medical image analysis.

Main Methods:

• Proposed a new approach optimizing the loss function using the optimal transport method.
• The optimal transport method rigorously calculates differences between predicted and annotated cell count maps.
• Evaluated the algorithm on three public datasets: VGG, MBM, and ADI.

Main Results:

• Achieved superior performance compared to state-of-the-art methods on all tested datasets.
• Reported mean absolute errors (MAE) of 2.3 (VGG), 4.8 (MBM), and 13.1 (ADI).
• Demonstrated smaller standard deviations, indicating higher reliability.

Conclusions:

• The optimal transport method offers a more accurate and reliable solution for cell counting in medical image processing.
• This advancement has significant implications for quantitative analysis in microscopic imaging.