Deep Active Learning Framework for Lymph Node Metastasis Prediction in Medical Support System.

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Histopathological analysis of sentinel axillary lymph nodes is crucial for breast cancer staging.
• Deep learning (DL) offers potential for auxiliary medical systems to aid pathologists.
• High annotation costs due to complex medical images hinder DL model development.

Purpose of the Study:

• To develop a cost-effective DL framework for breast cancer staging using limited labeled data.
• To improve diagnostic precision and accuracy in histopathological analysis of lymph nodes.
• To reduce the burden on pathologists by minimizing annotation requirements.

Main Methods:

• A novel DL framework with a three-stage query strategy and a unique model update strategy.
• An auto-encoder is trained on all samples for global representation.
• Sample selection involves uncertainty, coreset-based redundancy reduction, and distribution difference evaluation; model updates use frozen weights and a smaller learning rate.

Main Results:

• The proposed method achieves performance comparable to training with all labeled data, using only 50% of labeled samples.
• Iterative efficiency is improved compared to uncertainty, representative, or hybrid strategies on lymph node and other datasets.
• Model update strategies significantly reduce training costs compared to fine-tuning, retraining, or replay methods.

Conclusions:

• The developed DL framework effectively reduces annotation costs and training time for breast cancer staging.
• The three-stage query and novel model update strategies enable high performance with minimal labeled data.
• This approach offers a promising solution for developing accurate and efficient auxiliary diagnostic systems in computational pathology.