Microscopic Hematological Image Classification with Captions Using Few-Shot Learning in Data-Scarce Environments

Authors:

Abstract
Recent advancements in vision-language models (VLMs) have opened new possibilities for few-shot learning in medical image analysis. This study explores the application of VLMs for microscopic blood cell image classification using descriptive captions. We investigate the performance of open-source models like CLIP, PLIP, and BiomedCLIP on this task, focusing on prompt engineering techniques and their impact on classification accuracy. Our experiments show that Biomed-CLIP consistently outperforms other models in few-shot learning scenarios, delivering superior results using only 10% of the training data, compared to traditional CNN-based approaches that require 70% to 80% of the training data to achieve similar performance. While BiomedCLIP showed lower accuracy in zero-shot learning, it demonstrated substantial improvements in few-shot settings. By leveraging contextual prompts, our approach effectively classifies both normal and malignant blood cells across various categories. These findings suggest that few-shot learning with VLMs can mitigate data scarcity issues in hematological image analysis, providing a promising methodology for early diagnosis and treatment of hematological disorders.

This publication has 27 references indexed in Scilit: