Fine-tuning of advanced vision language models for peripheral blood cell image analysis

This thesis investigates the application of vision-language models (VLMs) for automated morphological analysis of peripheral blood cells. While manual microscopic analysis remains the gold standard in hematological diagnosis, it is time-consuming and subject to inter-observer variability. This work...

Descripción completa

Detalles Bibliográficos
Autor: Ruiz Martínez, Laura
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Universitat Oberta de Catalunya (UOC)
Repositorio:O2, repositorio institucional de la UOC
OAI Identifier:oai:openaccess.uoc.edu:10609/152020
Acceso en línea:https://hdl.handle.net/10609/152020
Access Level:acceso abierto
Palabra clave:machine learning
medical imaging analysis
deep learning
fine-tuning
peripheral blood
cell classification
visionlanguage models (VLM)
Bioinformatics -- TFM
Bioinformàtica -- TFM
Descripción
Sumario:This thesis investigates the application of vision-language models (VLMs) for automated morphological analysis of peripheral blood cells. While manual microscopic analysis remains the gold standard in hematological diagnosis, it is time-consuming and subject to inter-observer variability. This work aims to develop and evaluate fine-tuned VLMs capable of generating accurate morphological descriptions of blood cells from microscopic images. Our methodology comprised three main phases: First, we created a synthetic dataset of 10,000 peripheral blood cell images paired with expert-crafted morphological descriptions. Second, we implemented and compared fine-tuning approaches using Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA) on three open-source VLMs: LLAMA 3.2, QWEN, and SMOVLM. Finally, we developed a web-based interface for practical deployment. Results demonstrated significant improvements across all models post-fine-tuning, with QWEN achieving the highest performance (BLEU: 0.22, ROUGE-1: 0.55, BERTScore F1: 0.89). To ensure accessibility and enable ongoing evaluation, the model has been deployed as a web application on Hugging Face Spaces, making it freely available to the research community. We conclude that fine-tuned VLMs can effectively analyze peripheral blood cell morphology, offering potential for standardizing hematological analysis. This work establishes a framework for adapting vision-language models to specialized medical imaging tasks, with implications for improving diagnostic workflows in clinical settings. The complete implementation is available at Github.