Practical machine learning for disease diagnosis.

Area of Science:

• Microscopy
• Computational Biology
• Artificial Intelligence

Background:

• Deep learning neural networks (DNNs) are essential for modern microscopy analysis.
• Accurate annotation of cell images is a critical bottleneck for DNN training.
• Existing methods demand cell-level ground truth, which is labor-intensive.

Purpose of the Study:

• To streamline the annotation process for training deep learning models in microscopy.
• To reduce the dependency on precise cell-level annotations for disease classification.

Main Methods:

• Implementation of a novel network training strategy using patient-level disease classification.
• Leveraging existing patient data for model training, bypassing the need for cell-level annotation.

Main Results:

• Demonstrated feasibility of training deep learning models with patient-level data.
• Significantly reduced the annotation effort required for microscopy image analysis.

Conclusions:

• Patient-level classification offers an efficient alternative to cell-level annotation for deep learning in microscopy.
• This approach enhances the accessibility and scalability of DNNs for biological and medical imaging analysis.