medgemma-1.5-4b-it-leukemia-lora

🩸 MedGemma Leukemia LoRA - Blood Cell Classification Model

<p align="center"> <img src="https://img.shields.io/badge/Accuracy-78.15%25-brightgreen" alt="Accuracy"/> <img src="https://img.shields.io/badge/Leukemia%20Recall-83.10%25-orange" alt="Recall"/> <img src="https://img.shields.io/badge/F1%20Score-83.24%25-blue" alt="F1"/> <img src="https://img.shields.io/badge/Training%20Images-10%2C000+-purple" alt="Training"/> </p>

🎯 Model Summary

medgemma-1.5-4b-it-leukemia-lora is a parameter-efficient fine-tuned vision-language model specialized for Acute Lymphoblastic Leukemia (ALL) detection from microscopic blood cell images.

Built on Google's MedGemma 1.5 4B IT foundation model and fine-tuned using LoRA (Low-Rank Adaptation), this model achieves high recall for leukemia detection—critical for medical screening applications where minimizing false negatives saves lives.

💡 Key Impact: Early detection of leukemia improves 5-year survival rates from 20% to over 85%.

---

---

⚠️ CRITICAL DISCLAIMER - READ BEFORE USE

THIS MODEL IS FOR RESEARCH AND EDUCATIONAL PURPOSES ONLY

🚫 What This Model Is NOT:

• ❌ NOT FDA-approved or cleared for medical use
• ❌ NOT CE-marked for clinical diagnosis in EU
• ❌ NOT validated for real-world patient care
• ❌ NOT a substitute for professional medical diagnosis
• ❌ NOT suitable for making clinical treatment decisions

✅ What This Model IS:

• ✅ A research demonstration of LoRA fine-tuning techniques
• ✅ An educational tool for understanding medical AI
• ✅ A proof-of-concept for the MedGemma Impact Challenge
• ✅ A benchmark for comparing medical vision-language models

📊 Performance Metrics

🧬 Model Details

Architecture & Training

Developers

• Chaudhry Muhammad Suleman - @chaudhrysuleman
• Muhammad Idnan - @good2idnan

Training Data

The model was trained on the C-NMC Leukemia Classification Dataset, containing:

• Microscopic images of blood cells
• Binary labels: Normal (healthy) vs. ALL (Acute Lymphoblastic Leukemia)
• Single-cell centered images for consistent analysis

---

🚀 Quick Start

⚠️ Before Running This Code

WARNING: This code is for educational/research demonstration only.

• Do NOT use output for medical decisions
• Do NOT apply to real patient samples without proper validation and approval
• Ensure you have appropriate research ethics approval if using real medical data

---

Installation

pip install transformers peft torch pillow

Basic Usage

import os
import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForVision2Seq
from peft import PeftModel

BASE_MODEL_ID = "google/medgemma-1.5-4b-it"
LORA_ADAPTER_ID = "chaudhrysuleman/medgemma-1.5-4b-it-leukemia-lora"
hf_token = os.getenv("HF_TOKEN", "").strip() or None

processor = AutoProcessor.from_pretrained(
    BASE_MODEL_ID,
    trust_remote_code=True,
    token=hf_token
)

base_model = AutoModelForVision2Seq.from_pretrained(
    BASE_MODEL_ID,
    trust_remote_code=True,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto" if torch.cuda.is_available() else None,
    token=hf_token
)

model = PeftModel.from_pretrained(base_model, LORA_ADAPTER_ID, token=hf_token)
model.eval()

image = Image.open("blood_cell.jpg").convert("RGB")

prompt = (
    "Analyze this blood cell microscopy image and classify it.\n"
    "Is the cell NORMAL or LEUKEMIA (blast)?\n"
    "Answer with exactly one of: Normal, Leukemia."
)

messages = [{
    "role": "user",
    "content": [{"type": "image"}, {"type": "text", "text": prompt}]
}]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(images=image, text=text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=10, do_sample=False)

print(processor.decode(outputs[0], skip_special_tokens=True))

Try the Demo

🎮 Live Demo: LeukemiaScope on HuggingFace Spaces

---

💡 Intended Uses

✅ Appropriate Uses (Research & Education Only)

Required Context: All uses must be in non-clinical, research, or educational settings with proper supervision and understanding of limitations.

---

🚫 Prohibited Uses

Legal Requirement: Any clinical application requires separate regulatory approval in your jurisdiction (FDA 510(k), CE-IVD marking, etc.).

⚠️ Limitations & Risks

Known Limitations

1. Dataset Bias: Trained on a single public dataset (C-NMC); may not generalize to:

   • Different microscope types
   • Alternative staining techniques
   • Other leukemia subtypes (AML, CML, CLL)
   • Varying image resolutions

2. Population Representation: Dataset demographics may not reflect all patient populations

3. False Negatives/Positives: While optimized for high recall, some cases may be misclassified

Safety Recommendations

• ✅ Always validate on local clinical data before any experimental use
• ✅ Use as screening support, not final diagnosis
• ✅ Combine with expert pathologist review
• ✅ Follow regulatory guidelines for medical AI in your jurisdiction
• ❌ Never deploy in clinical settings without proper approval

---

🔬 Technical Specifications

Input Requirements

• Image Format: RGB images (PNG, JPG, JPEG)
• Recommended: Single blood cell, centered, clear microscopy
• Resolution: Model handles various resolutions (224x224+ recommended)

Compute Requirements

---

📚 Citation

If you use this model in your research, please cite:

@misc{medgemma-leukemia-lora-2026,
  author = {Suleman, Chaudhry Muhammad and Idnan, Muhammad},
  title = {MedGemma Leukemia LoRA: Parameter-Efficient Fine-Tuning for Blood Cell Classification},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/chaudhrysuleman/medgemma-1.5-4b-it-leukemia-lora}
}

---

🔗 Resources

---

📜 License

This model is released under the Apache 2.0 License, consistent with the MedGemma base model license. See the LICENSE file for details.

---

🔒 Responsible AI Statement

We are committed to the responsible development and use of AI in healthcare:

• Transparency: All limitations and risks are openly disclosed
• Safety: Strong disclaimers against clinical misuse
• Accountability: Contact information provided for reporting concerns
• Education: Clear guidance on appropriate vs. prohibited uses
• Compliance: Encouragement to follow all applicable regulations

Version History

• v1.0 (February 2026): Initial release for MedGemma Impact Challenge
  • 83.10% recall on C-NMC test set
  • Research/educational use only

---

Remember: This is a research tool, not a medical device. Human expertise saves lives, not AI alone. 🏥

Built with ❤️ for the MedGemma Impact Challenge 2026