Generalising from conventional pipelines using deep learning in high-throughput screening workflows.

Area of Science:

• Computational Biology
• Biomedical Imaging
• Machine Learning

Background:

• Precision medicine requires large datasets for model development, often acquired through high-throughput screening.
• Accurate image analysis is critical for the quality of high-throughput screening results.
• Manual ground truth labeling for deep learning in image segmentation is time-consuming and expensive for experimental labs.

Purpose of the Study:

• To develop a deep learning model for image segmentation that utilizes weakly labeled data.
• To reduce the reliance on manually annotated datasets for training machine learning models in microscopy.
• To improve the efficiency and accuracy of image analysis in biological research.

Main Methods:

• A deep learning network was trained using weak training labels automatically generated by conventional computer vision methods.
• The developed solution was integrated into a user-friendly graphical interface for prediction assessment and correction.
• The approach was compared against traditional methods using a case study on autophagy event segmentation.

Main Results:

• The deep learning network achieved a 25% increase in mean intersection over union compared to conventional methods.
• The proposed method significantly reduced both development and inference times.
• The system demonstrated superior generalization beyond noisy labels, outperforming models trained on small, manually curated datasets.

Conclusions:

• Combining deep learning with automatically generated weak labels offers a cost-effective and efficient alternative for image segmentation in microscopy.
• The hybrid approach leverages computer strengths in pixel-level segmentation and human expertise in contextual interpretation.
• This work facilitates the translation of advanced imaging technologies into practical laboratory applications for complex disease research.