HistoNeXt: dual-mechanism feature pyramid network for cell nuclear segmentation and classification

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Digital pathology enables automated analysis of histological images.
• Accurate nuclear segmentation and classification are crucial for disease diagnosis.
• Existing models face challenges in efficiency and accuracy for complex tasks.

Purpose of the Study:

• To develop an end-to-end convolutional neural network (CNN) for analyzing H&E-stained histological images.
• To enhance the performance and efficiency of nuclear segmentation and classification.
• To integrate these tasks within a digital pathology workflow.

Main Methods:

• Developed HistoNeXt, an encoder-decoder CNN model using ConvNeXt framework.
• Implemented a dual-mechanism feature pyramid fusion for segmentation and classification.
• Utilized densely connected blocks and channel attention for feature extraction and refinement.
• Applied extensive data augmentation and type-aware sampling to address class imbalance.

Main Results:

• HistoNeXt achieved competitive performance on multiple public datasets (CONSEP, PanNuke, CPM17, KUMAR).
• Demonstrated high scores in Dice Similarity Coefficient (DICE), Aggregated Jaccard Index (AJI), and Panoptic Quality (PQ).
• Achieved overall F1 scores up to 0.82 and cell-type specific F1 scores, with low computational complexity (33.7 GFLOPS).

Conclusions:

• HistoNeXt enhances precision and efficiency in nuclear segmentation and classification through novel CNN architecture.
• Low computational complexity allows for deployment in resource-constrained environments.
• Represents a significant advancement for CNN applications in digital pathology analysis.