Towards pixel-to-pixel deep nucleus detection in microscopy images

Area of Science:

• Biomedical image analysis
• Computational pathology
• Machine learning in microscopy

Background:

• Nucleus detection is crucial for quantitative microscopy studies.
• Deep neural networks, particularly CNNs, are powerful for nucleus detection.
• Existing models often lack generalizability across diverse microscopy datasets.

Purpose of the Study:

• To systematically analyze the applicability of deep learning models for nucleus detection across varied microscopy image data.
• To address critical, understudied questions regarding model generalizability in nucleus detection.

Main Methods:

• Developed and evaluated a fully convolutional network-based regression model.
• Extensively tested the model on large-scale digital pathology and microscopy datasets from multiple institutions and 23 organ types.
• Investigated the impact of training data origin and composition on model performance.

Main Results:

• Nucleus detection models often require training data from the same organ types as the target dataset.
• Models trained on different organ types may perform poorly, even with similar imaging conditions, requiring fine-tuning.
• Training with mixed data does not guarantee improved accuracy; careful data manipulation is key.

Conclusions:

• Conducted a systematic case study on deep models for nucleus detection in diverse microscopy images.
• Presented an end-to-end, pixel-to-pixel fully convolutional regression network.
• Reported significant findings on model generalizability and performance, aiding future nucleus detection research.