NuHTC: A hybrid task cascade for nuclei instance segmentation and classification.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Accurate nuclei instance segmentation and classification are crucial for cancer diagnosis and prognosis.
• Existing methods often use single-level features, which are less effective for diverse nuclei.
• Multilevel feature maps are better suited for handling variations in nuclei size and type.

Purpose of the Study:

• To develop an effective top-down nuclei instance segmentation and classification framework (NuHTC).
• To improve the precision of bounding box prediction and feature learning for nuclei instances.
• To enhance the performance of digital pathology image analysis for cancer-related tasks.

Main Methods:

• Developed a hybrid task cascade (HTC) based framework named NuHTC.
• Introduced a watershed proposal network (WSPN) to enhance region proposal network accuracy.
• Integrated a hybrid feature extractor (HFE) for improved utilization of global and semantic features at the RoI alignment stage.

Main Results:

• NuHTC demonstrated superior performance in both instance segmentation and classification tasks.
• Extensive experiments were conducted on four public multiclass nuclei instance segmentation datasets.
• The proposed method showed significant improvements compared to existing state-of-the-art approaches.

Conclusions:

• NuHTC effectively addresses limitations of previous nuclei segmentation methods.
• The framework enhances the learning of nuclei instance features with reduced intraclass variance.
• NuHTC offers a promising advancement for automated cancer diagnosis and prognosis using digital pathology.