A CLIP-based framework for multiclass lung histopathology classification with prompt engineering and...

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate lung cancer histopathological classification is critical for patient diagnosis and treatment.
• Existing methods may face challenges in distinguishing subtle differences between lung cancer subtypes.
• Automated classification systems can enhance efficiency and consistency in pathology workflows.

Purpose of the Study:

• To develop and evaluate a Contrastive Language Image Pretraining (CLIP)-based framework for multiclass lung histopathology classification.
• To differentiate between benign lung tissue, lung adenocarcinoma, and lung squamous cell carcinoma using a multimodal approach.
• To assess the performance and robustness of the proposed CLIP framework in automated lung cancer diagnosis.

Main Methods:

• Utilized a pretrained CLIP ViT-B/32 backbone for feature extraction.
• Employed domain-specific prompt engineering and multimodal image-text pairing.
• Implemented similarity-based classification within a shared embedding space.
• Incorporated data augmentation, Focal Loss, AdamW optimization, and learning rate scheduling for robust fine-tuning.

Main Results:

• Achieved a best validation accuracy of 95.20% on lung histopathology classification.
• Obtained a macro Area Under the Curve (AUC) of 0.9870 and a micro AUC of 0.9877.
• Demonstrated steady performance improvement across training epochs, with early stopping at epoch 23.

Conclusions:

• Integrating CLIP with pathology-specific text prompts offers a reliable framework for automated lung cancer histopathology classification.
• The proposed method shows significant potential for advancing intelligent digital pathology systems.
• This approach can aid in timely diagnosis and treatment planning for lung cancer patients.