Assessing CNNs and LoRA-Fine-Tuned Vision-Language Models for Breast Cancer Histopathology Image Classification.

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Histopathology classification of breast cancer is challenging due to morphological variations across magnifications.
• Convolutional Neural Networks (CNNs) are standard, but Vision-Language Models (VLMs) offer potential for complex medical images.

Purpose of the Study:

• To systematically compare the performance of fine-tuned VLMs (Qwen2, SmolVLM) against CNN baselines for breast cancer histopathology classification.
• To evaluate model performance across various magnifications (40× to 400×).

Main Methods:

• Utilized the BreakHis dataset for training and evaluation.
• Fine-tuned VLMs (Qwen2, SmolVLM) using Low-Rank Adaptation (LoRA).
• Evaluated models using accuracy, precision, recall, F1-score, and AUC across magnifications.

Main Results:

• CNN baselines (ResNet34) achieved the highest performance across all magnifications.
• SmolVLM demonstrated competitive performance, narrowing the gap with CNNs, especially at 200×.
• Qwen2 showed moderate performance, consistently outperformed by SmolVLM.

Conclusions:

• LoRA-fine-tuned VLMs offer parameter-efficient alternatives for digital pathology, performing competitively with CNNs.
• CNNs currently provide superior accuracy and robustness for histopathology classification.
• VLMs show potential for resource-constrained settings in computational pathology.