FKDNuSeg: Flawless knowledge distillation for lightweight and fast nuclei instance segmentation and classification.

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Nuclei segmentation and classification are crucial in computational pathology.
• Current methods face efficiency challenges with large whole slide images and complex models, limiting clinical application.
• There is a need for efficient and accurate nuclei analysis tools in pathology.

Purpose of the Study:

• To develop a lightweight and fast nuclei segmentation and classification model named FKDNuSeg.
• To address the efficiency limitations of existing computational pathology models.
• To improve the viability of automated nuclei analysis in clinical settings.

Main Methods:

• FKDNuSeg utilizes a knowledge distillation strategy within a multi-task learning architecture, employing ENet as the backbone.
• A novel knowledge distillation approach ensures the student model learns effectively from the teacher model, mitigating bias from flawed data, especially for minority classes.
• An edge detection task, enhanced by a curvature module, refines nuclei edge extraction and separation of overlapping nuclei.

Main Results:

• FKDNuSeg is over 10x smaller than state-of-the-art models, with only 4.11M parameters.
• The model achieves a fast patch-level inference time of 1.38 ms on an RTX 3090 GPU.
• Experiments on CoNSeP and PanNuke datasets confirm reduced model complexity and improved efficiency with maintained performance.

Conclusions:

• FKDNuSeg offers a significant advancement in efficient nuclei segmentation and classification for computational pathology.
• The model's lightweight design and high speed make it suitable for real-world clinical applications.
• FKDNuSeg demonstrates that efficient deep learning models can achieve considerable performance in complex medical image analysis tasks.