AMLPF-CLIP: Adaptive Prompting and Distilled Learning for Imbalanced Histopathological Image Classification.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Histopathological image classification (HIC) is crucial for computer-aided diagnosis, aiding in tasks like tumor grading and survival prediction.
• Current HIC methods struggle with integrating domain knowledge, handling imbalanced datasets, and maintaining computational efficiency.

Purpose of the Study:

• To introduce AMLPF-CLIP, an enhanced CLIP-based framework designed to overcome the limitations of existing HIC methods.
• To improve semantic representation, cross-modal alignment, and classification performance in HIC.

Main Methods:

• Developed an Adaptive Multi-Level Prompt Fusion (AMLPF) strategy using class labels, descriptions, and GPT-4o-generated pathological features.
• Implemented a class-balanced resampling method to address data imbalance and improve performance on underrepresented classes.
• Utilized Knowledge Distillation (KD) with L2 loss to transfer knowledge from a large Vision Transformer (ViT-L/16) to a lightweight ResNet-50-based CLIP model.

Main Results:

• AMLPF-CLIP outperformed eleven state-of-the-art methods on three public datasets (Chaoyang, BreaKHis, LungHist700).
• Achieved accuracy improvements of 1.19% on Chaoyang, 2.64% on BreaKHis, and 0.90% on LungHist700.
• Demonstrated enhanced robustness and computational efficiency compared to existing approaches.

Conclusions:

• AMLPF-CLIP offers a significant advancement in HIC by effectively integrating domain knowledge and addressing class imbalance.
• The framework shows practical applicability due to its improved accuracy, robustness, and efficiency in computer-aided diagnosis.