CytoAL: Towards Label-Efficient Cytology Diagnosis via Cellularity-Guided Active Learning

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Cytopathology

Background:

• Thyroid fine needle aspiration (FNA) is crucial for cancer risk stratification and treatment decisions.
• Current AI diagnostic strategies for thyroid cytology are hindered by dispersed diagnostic cues and labor-intensive annotation.
• Limited public datasets for thyroid cytology pose a challenge for developing robust AI models.

Purpose of the Study:

• To develop an efficient active learning framework (Cyto-AL) for thyroid cytology image analysis.
• To improve the prediction accuracy of The Bethesda System for Reporting Thyroid Cytology (TBSRTC) categories using Whole Slide Images (WSIs).
• To create a benchmark dataset for thyroid cytology to facilitate AI research.

Main Methods:

• Proposed a cellularity-based active learning framework (Cyto-AL) correlating cellularity with diagnostic categories for query optimization.
• Implemented a severe-stage pinpointed Multiple Instance Learning (MIL) approach for WSI TBSRTC prediction.
• Developed a lightweight score model to optimize human-in-the-loop (HITL) annotation efficiency.

Main Results:

• Cyto-AL achieved a 2.2% average classification accuracy improvement over state-of-the-art methods with equivalent labeled data.
• The lightweight ranking-aware model reduced training time by approximately 65%.
• The WSI-level MIL approach improved average accuracy by 10.7% and Macro-F1 score by 3.5% compared to standard sampling methods.

Conclusions:

• The proposed Cyto-AL framework significantly enhances AI-driven thyroid cytology analysis, improving diagnostic accuracy and annotation efficiency.
• The novel MIL approach offers superior performance for WSI TBSRTC classification.
• The released benchmark dataset and source code will accelerate research in AI for thyroid cancer diagnosis.