VLM-CPL: Consensus Pseudo-Labels From Vision-Language Models for Annotation-Free Pathological Image Classification.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computational Pathology

Background:

• Deep learning for cancer diagnosis requires extensive labeled pathological images, posing a significant annotation burden.
• Existing methods struggle with noisy labels inherent in automated approaches.

Purpose of the Study:

• To develop a human annotation-free method for pathological image classification.
• To leverage Vision-Language Models (VLMs) for accurate cancer diagnosis without manual labeling.

Main Methods:

• Utilized zero-shot inference from pre-trained VLMs to generate initial pseudo-labels.
• Introduced VLM-CPL, incorporating two noisy label filtering techniques and semi-supervised learning.
• Developed prompt-based and feature-based pseudo-labeling with uncertainty estimation and clustering.
• Implemented prompt-feature consensus and High-confidence Cross Supervision for reliable learning.
• Employed an open-set prompting strategy to refine patch selection from whole slides.

Main Results:

• The proposed method significantly outperformed direct zero-shot VLM classification on pathological images.
• Achieved superior performance compared to existing noisy label learning techniques.
• Demonstrated effectiveness across five public pathological image datasets for both patch-level and slide-level classification.

Conclusions:

• The VLM-CPL approach offers a powerful, annotation-free solution for pathological image classification.
• Leveraging VLMs with advanced label-filtering strategies enhances automated cancer diagnosis.
• The method shows promise for reducing the reliance on manual annotation in digital pathology.