The paper proposes a novel Triple U-Net architecture for nuclei instance segmentation in cryosectioned H&E stained histological images. The key highlights are:
The Triple U-Net model consists of three branches - RGB, Hematoxylin, and Segmentation. The Hematoxylin branch extracts contour-aware features to enhance the segmentation accuracy, while the Segmentation branch fuses the features from the other two branches.
A Progressive Dense Feature Aggregation (PDFA) module is introduced to effectively combine features from the different branches, improving feature representation learning.
Watershed post-processing is applied to the segmentation maps to further refine the instance segmentation results.
Extensive experiments are conducted on the CryoNuSeg dataset, a novel fully annotated dataset of cryosectioned H&E stained nuclei images. The proposed Triple U-Net architecture outperforms the baseline U-Net model, achieving an AJI score of 67.41% and a PQ score of 50.56%, which are significant improvements over the benchmark scores of 52.5% and 47.7%, respectively.
The model's performance is particularly strong on the AJI metric, which is a more strict evaluation measure that penalizes wrong predictions more than it rewards correct ones. This demonstrates the superior ability of the Triple U-Net to produce precise nuclei boundaries.
The proposed approach does not require any color normalization, making it more computationally efficient compared to other methods.
Overall, the Triple U-Net architecture represents a state-of-the-art solution for high-precision nuclei instance segmentation in cryosectioned histological images, with potential applications in rapid cancer diagnosis and informed surgical decision-making.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Zarif Ahmed,... at arxiv.org 04-22-2024
https://arxiv.org/pdf/2404.12986.pdfDeeper Inquiries