Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

HemBLIP: A Vision-Language Model for Interpretable Leukemia Cell Morphology Analysis

Julie van Logtestijn, Petru Manescu

TL;DR

Leukemia diagnosis from blood film morphology often suffers from black-box AI tools lacking interpretability. HemBLIP applies a vision-language framework to generate morphology-aware captions for healthy and leukemic white blood cells, linking textual outputs to diagnostic attributes. By constructing a morphology-rich ~14k image-caption dataset and comparing LoRA-based and full fine-tuning against MedGEMMA, HemBLIP achieves strong caption quality and attribute fidelity, with LoRA offering substantial efficiency gains. External validation demonstrates robust semantic alignment under domain shift, supporting the potential for transparent, scalable AI-assisted hematology diagnostics.

Abstract

Microscopic evaluation of white blood cell morphology is central to leukemia diagnosis, yet current deep learning models often act as black boxes, limiting clinical trust and adoption. We introduce HemBLIP, a vision language model designed to generate interpretable, morphology aware descriptions of peripheral blood cells. Using a newly constructed dataset of 14k healthy and leukemic cells paired with expert-derived attribute captions, we adapt a general-purpose VLM via both full fine-tuning and LoRA based parameter efficient training, and benchmark against the biomedical foundation model MedGEMMA. HemBLIP achieves higher caption quality and morphological accuracy, while LoRA adaptation provides further gains with significantly reduced computational cost. These results highlight the promise of vision language models for transparent and scalable hematological diagnostics.

HemBLIP: A Vision-Language Model for Interpretable Leukemia Cell Morphology Analysis

TL;DR

Leukemia diagnosis from blood film morphology often suffers from black-box AI tools lacking interpretability. HemBLIP applies a vision-language framework to generate morphology-aware captions for healthy and leukemic white blood cells, linking textual outputs to diagnostic attributes. By constructing a morphology-rich ~14k image-caption dataset and comparing LoRA-based and full fine-tuning against MedGEMMA, HemBLIP achieves strong caption quality and attribute fidelity, with LoRA offering substantial efficiency gains. External validation demonstrates robust semantic alignment under domain shift, supporting the potential for transparent, scalable AI-assisted hematology diagnostics.

Abstract

Microscopic evaluation of white blood cell morphology is central to leukemia diagnosis, yet current deep learning models often act as black boxes, limiting clinical trust and adoption. We introduce HemBLIP, a vision language model designed to generate interpretable, morphology aware descriptions of peripheral blood cells. Using a newly constructed dataset of 14k healthy and leukemic cells paired with expert-derived attribute captions, we adapt a general-purpose VLM via both full fine-tuning and LoRA based parameter efficient training, and benchmark against the biomedical foundation model MedGEMMA. HemBLIP achieves higher caption quality and morphological accuracy, while LoRA adaptation provides further gains with significantly reduced computational cost. These results highlight the promise of vision language models for transparent and scalable hematological diagnostics.
Paper Structure (13 sections, 2 figures, 3 tables)

This paper contains 13 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Dataset Construction
  4. Model Architectures and Fine-tuning
  5. Clinical Relevance Evaluation
  6. External Validation
  7. Results
  8. Caption Generation Performance
  9. Morphological Attribute Accuracy
  10. HemBLIP as a feature extractor
  11. Conclusion
  12. Acknowledgments
  13. Compliance with Ethical Standards

Figures (2)

  • Figure 1: Comparison of the two vision--language architectures used in this study. Both were fine-tuned on the morphology-aware leukemia dataset using LoRA adapters for parameter-efficient adaptation.
  • Figure 2: Example blood cell inputs with ground-truth captions and model-generated descriptions. Each model exhibits varying degrees of morphological specificity and diagnostic accuracy.