Automated Training of Deep Convolutional Neural Networks for Cell Segmentation.

Area of Science:

• Computer Vision
• Biomedical Imaging
• Machine Learning

Background:

• Deep Convolutional Neural Networks (DCNNs) are advanced algorithms for image segmentation.
• High-performance DCNNs require extensive, problem-specific training datasets.
• Manual annotation of training data is time-consuming and labor-intensive.

Purpose of the Study:

• To investigate the efficacy of automatically generated ground truth for training DCNNs.
• To compare the performance of DCNNs trained on automated versus manual annotations.
• To reduce the burden of manual data labeling in DCNN-based image segmentation.

Main Methods:

• Utilized fluorescently labeled cells to create automated ground truth data.
• Trained Deep Convolutional Neural Networks using this automatically generated dataset.
• Evaluated DCNN performance against models trained with manually annotated data.

Main Results:

• DCNNs trained with automatically generated ground truth demonstrated performance comparable to those trained with manual annotations.
• Automated annotation significantly reduced the time and resources required for dataset creation.
• The findings suggest a viable alternative to manual labeling for DCNN training.

Conclusions:

• Automatic ground truth generation using fluorescently labeled cells is an effective strategy for training DCNNs in image segmentation.
• This approach offers a scalable and efficient method for developing high-performing DCNN models.
• Reduces dependency on manual annotation, accelerating research and application development.