NuClick: A deep learning framework for interactive segmentation of microscopic images.

Area of Science:

• Computational Pathology
• Medical Imaging Analysis
• Deep Learning

Background:

• Accurate object segmentation is crucial for computational pathology workflows.
• Deep learning models require extensive labeled data, which is costly and time-consuming to acquire in medical imaging.
• Expert annotation is essential but labor-intensive for nuclei, cells, and glands.

Purpose of the Study:

• To introduce NuClick, a CNN-based approach for efficient annotation of cellular and glandular structures in pathology images.
• To reduce the annotation effort required from human experts while maintaining high precision.
• To enable faster development and deployment of computational pathology tools.

Main Methods:

• NuClick utilizes a Convolutional Neural Network (CNN) architecture.
• For individual objects like nuclei and cells, a single click annotation is sufficient.
• For complex structures like glands, a 'squiggle' input guides the segmentation.
• Annotation signals are incorporated as auxiliary inputs alongside RGB image data.

Main Results:

• NuClick achieves precise segmentation with minimal user interaction (one click for cells/nuclei, squiggle for glands).
• The method demonstrates robustness across various object scales and input variations.
• NuClick is adaptable to new imaging domains and yields reliable annotations.
• An instance segmentation model trained with NuClick annotations achieved top performance in the LYON19 challenge.

Conclusions:

• NuClick offers a highly efficient and accurate solution for object annotation in computational pathology.
• The approach significantly lowers the barrier to creating large, high-quality datasets for deep learning in medical imaging.
• NuClick facilitates downstream analysis and improves the performance of pathology-related AI models.