Improving lung cancer pathological hyperspectral diagnosis through cell-level annotation refinement.

Area of Science:

• Pathology
• Medical Imaging
• Computational Biology

Background:

• Lung cancer diagnosis relies heavily on accurate pathological examination.
• Current hyperspectral lung cancer datasets suffer from coarse manual annotations, hindering deep learning model performance.
• Precise cell-level annotations are essential for advanced pathological image analysis.

Purpose of the Study:

• To enhance hyperspectral pathological image analysis for lung cancer.
• To develop a semi-automated method for refining cell-level annotations in hyperspectral lung tumor datasets.
• To create a high-quality hyperspectral dataset for improved pathological diagnosis.

Main Methods:

• Employed K-means unsupervised clustering combined with human-guided selection for annotation refinement.
• Utilized spectral features for accurate cell-level mask generation.
• Validated the method on a hyperspectral lung squamous cell carcinoma dataset (65 samples).

Main Results:

• Improved pixel-level segmentation accuracy from 77.33% to 92.52%.
• Significantly reduced annotation time per slide from over 30 minutes to approximately 5 minutes.
• Decreased prediction noise in segmentation results.

Conclusions:

• The semi-automated annotation refinement method effectively addresses challenges in hyperspectral pathological analysis.
• The approach enhances the accuracy and efficiency of creating high-quality hyperspectral datasets for lung cancer research.
• Improved annotations facilitate better deep learning model training and pathological diagnosis.