DeepDistance: A multi-task deep regression model for cell detection in inverted microscopy images.

Area of Science:

• Computational biology
• Image analysis
• Deep learning

Background:

• Accurate cell detection and segmentation are crucial for biological research.
• Existing regression-based methods have limitations in handling complex cell imaging data.

Purpose of the Study:

• To introduce DeepDistance, a novel deep regression model for cell detection in inverted microscopy images.
• To enhance cell localization and boundary delineation using a multi-task learning framework.

Main Methods:

• Developed DeepDistance, a fully convolutional network (FCN) employing multi-task regression.
• Incorporated complementary tasks: learning an inner distance metric and a normalized outer distance.
• Introduced an extended version with an auxiliary classification task for improved performance.

Main Results:

• DeepDistance successfully identifies cell locations and delineates cell boundaries across multiple human cell lines.
• The model demonstrated robust performance, including on a cell line not used during training.
• Multi-task learning with shared feature representations significantly improved detection and segmentation accuracy.

Conclusions:

• The proposed DeepDistance model and its extended version offer a powerful solution for automated cell analysis in microscopy.
• Multi-task learning is an effective strategy for improving cell detection and segmentation accuracy.
• DeepDistance provides a significant advancement over existing regression-based methods.