A novel expert-annotated single-cell dataset for thyroid cancer diagnosis with deep learning benchmarks

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Automated cytological analysis is crucial for accurate cancer diagnosis.
• High-quality, annotated datasets are essential for developing reliable AI diagnostic tools.
• Thyroid cancer diagnosis benefits from detailed analysis of nuclear features in single cells.

Purpose of the Study:

• To introduce a novel, expert-annotated single-cell image dataset for thyroid cancer diagnosis.
• To establish deep learning baseline models for multi-label classification on this dataset.
• To provide a benchmark for future research in AI-based cytological diagnosis.

Main Methods:

• A dataset of 3,419 single-cell images from histopathological slides was created and annotated with nine nuclear features.
• Deep learning pipelines using ConvNeXt, Vision Transformers (ViT), and ResNet were developed.
• Techniques for class imbalance, including conditional CutMix, weighted sampling, and SPA loss with Label Pairwise Regularization (LPR), were implemented.

Main Results:

• The proposed deep learning pipelines demonstrated good performance on the dataset.
• The study highlighted the dataset's characteristics as challenging for AI models.
• The results confirmed the effectiveness of specific model architectures and data-centric strategies.

Conclusions:

• The expert-annotated dataset is a valuable resource for advancing automated cytological analysis in thyroid cancer.
• Deep learning models, particularly with effective architectures and data strategies, show potential for accurate multi-label classification.
• This work sets a benchmark for interpretable and reliable AI-based cytological diagnosis.