Mixed-Supervised Learning for Cell Classification.

Area of Science:

• Computational biology
• Medical imaging analysis
• Artificial intelligence in medicine

Background:

• Accurate cell classification from histopathology images is vital for cancer diagnosis and tumor recognition.
• Deep learning significantly improves classification accuracy, with semi-supervised learning leveraging both labeled and unlabeled data.
• Complex datasets can lead to models learning detrimental features, necessitating human oversight in training.

Purpose of the Study:

• To develop a mixed-supervised method integrating semi-supervision and human-in-the-loop strategies for enhanced cell classification.
• To improve the robustness and accuracy of deep learning models in histopathology image analysis.
• To address the challenge of learning harmful features in complex, diverse datasets.

Main Methods:

• A novel mixed-supervised approach combining semi-supervised learning with human-in-the-loop guidance was proposed.
• A sample selection mechanism was designed to differentiate between confident unlabeled samples for automatic optimization and unreliable ones for human correction.
• The model was pre-trained using prior human annotations and fine-tuned with pseudo labels and online human annotations.

Main Results:

• The mixed-supervised model achieved high overall accuracies: 86.56% on LUSC, 99.33% on BloodCell, and 74.12% on PanNuke datasets.
• The integration of human guidance effectively mitigated the learning of harmful features in complex datasets.
• The method demonstrated superior performance compared to traditional semi-supervised approaches.

Conclusions:

• The proposed mixed-supervised method offers a powerful strategy for accurate cell classification in histopathology.
• Incorporating human-in-the-loop feedback significantly enhances deep learning model performance in medical image analysis.
• This approach holds promise for improving diagnostic accuracy in oncology and related fields.