flare25-qwen2.5vl

QWen2.5VL Fine-tuned for FLARE 2025 Medical Image Analysis

This model is a fine-tuned version of Qwen/Qwen2.5-VL-7B-Instruct specifically optimized for medical image analysis tasks in the FLARE 2025 2D Medical Multimodal Dataset challenge.

Model Description

• Base Model: Qwen2.5-VL-7B-Instruct
• Fine-tuning Method: QLoRA (Low-Rank Adaptation)
• Target Domain: Medical imaging across 8 modalities (Clinical, Dermatology, Endoscopy, Mammography, Microscopy, Retinography, Ultrasound, Xray)
• Tasks: Medical image captioning, visual question answering, report generation
• Training Data: 19 FLARE 2025 datasets with comprehensive medical annotations

Training Details

Training Data

The model was fine-tuned on 19 diverse medical imaging datasets from FLARE 2025, details can be found at: https://huggingface.co/datasets/FLARE-MedFM/FLARE-Task5-MLLM-2D

Training Configuration

# LoRA Configuration
lora_r: 16
lora_alpha: 32
lora_dropout: 0.1
target_modules: ['k_proj', 'v_proj', 'o_proj', 'gate_proj', 'down_proj', 'up_proj', 'q_proj']
task_type: CAUSAL_LM

# Training Statistics
total_steps: 1000
learning_rate: N/A
final_eval_loss: 5.4849

Training Procedure

• Optimization: 4-bit quantization with BitsAndBytesConfig
• LoRA Configuration:
  • r=64, alpha=16, dropout=0.1
  • Target modules: All linear layers
• Memory Optimization: Gradient checkpointing, flash attention
• Batch Size: Dynamic based on image resolution
• Learning Rate: 1e-4 with cosine scheduling
• Training Steps: 1000 steps with evaluation every 500 steps

Model Performance

This model has been evaluated across multiple medical imaging tasks with the following capabilities:

• Image Captioning: Generates detailed medical reports from imaging studies
• Visual Question Answering: Answers clinical questions about medical images
• Diagnosis Support: Identifies pathological findings and abnormalities
• Multi-modal Understanding: Integrates visual and textual medical information

Evaluation Metrics

The model is evaluated using task-specific metrics following FLARE 2025 specifications:

Classification Tasks:

• Balanced Accuracy (PRIMARY): Handles class imbalance in medical diagnosis
• Accuracy: Standard classification accuracy
• F1 Score: Weighted F1 for multi-class scenarios

Multi-label Classification:

• F1 Score (PRIMARY): Sample-wise F1 across multiple labels
• Precision: Label prediction precision
• Recall: Label coverage recall

Detection Tasks:

• F1 Score @ IoU > 0.5 (PRIMARY): Standard computer vision detection metric
• Precision: Detection precision at IoU threshold
• Recall: Detection recall at IoU threshold

Instance Detection (Identity-Aware):

• F1 Score @ IoU > 0.3 (PRIMARY): Medical imaging standard for chromosome detection
• F1 Score @ IoU > 0.5: Computer vision standard
• Average F1: COCO-style average across IoU thresholds (0.3-0.7)
• Per-chromosome metrics: Detailed breakdown by chromosome identity

Counting Tasks:

• Mean Absolute Error (PRIMARY): Cell counting accuracy
• Root Mean Squared Error: Additional counting precision metric

Regression Tasks:

• Mean Absolute Error (PRIMARY): Continuous value prediction accuracy
• Root Mean Squared Error: Regression precision metric

Report Generation:

• GREEN Score (PRIMARY): Comprehensive medical report evaluation with 7 components:
  • Entity matching with severity assessment (30%)
  • Location accuracy with laterality (20%)
  • Negation and uncertainty handling (15%)
  • Temporal accuracy (10%)
  • Size/measurement accuracy (10%)
  • Clinical significance weighting (10%)
  • Report structure completeness (5%)
• BLEU Score: Text generation quality
• Clinical Efficacy: Medical relevance scoring

Usage

Installation

pip install transformers torch peft accelerate bitsandbytes

Basic Usage

import torch
from transformers import AutoTokenizer, AutoProcessor
from peft import PeftModel, PeftConfig
from PIL import Image

# Load the fine-tuned model
base_model_name = "Qwen/Qwen2.5-VL-7B-Instruct"
adapter_model_name = "leoyinn/qwen2.5vl-flare2025"

# Load tokenizer and processor
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
processor = AutoProcessor.from_pretrained(base_model_name)

# Load base model and adapter
from transformers import AutoModelForVision2Seq
base_model = AutoModelForVision2Seq.from_pretrained(
    base_model_name,
    torch_dtype=torch.float16,
    device_map="auto",
    load_in_4bit=True
)

# Load the fine-tuned adapter
model = PeftModel.from_pretrained(base_model, adapter_model_name)

# Prepare input
image = Image.open("medical_image.jpg")
prompt = "Describe the medical findings in this image."

# Process and generate
inputs = processor(
    images=image,
    text=prompt,
    return_tensors="pt"
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=True,
        temperature=0.7,
        top_p=0.9
    )

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

Limitations and Ethical Considerations

Limitations

• Model outputs may contain inaccuracies and should be verified by medical professionals
• Performance may vary across different medical imaging modalities and populations
• Training data may contain biases present in medical literature and datasets
• Model has not been validated in clinical settings

Intended Use

• Medical education and training
• Research in medical AI and computer vision
• Development of clinical decision support tools (with proper validation)
• Academic research in multimodal medical AI

Out-of-Scope Use

• Direct clinical diagnosis without physician oversight
• Treatment recommendations without medical professional validation
• Use in emergency medical situations
• Deployment in production clinical systems without extensive validation

Citation

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

@misc{qwen25vl-flare2025,
  title={QWen2.5VL Fine-tuned for FLARE 2025 Medical Image Analysis},
  author={Shuolin Yin},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/leoyinn/qwen2.5vl-flare2025}
}

@misc{qwen25vl-base,
  title={Qwen2.5-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution},
  author={Qwen Team},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct}
}

Model Details

• Model Type: Vision-Language Model (VLM)
• Architecture: QWen2.5VL with LoRA adapters
• Parameters: ~7B base parameters + LoRA adapters
• Precision: 4-bit quantized base model + full precision adapters
• Framework: PyTorch, Transformers, PEFT

Contact

For questions or issues, please open an issue in the model repository or contact the authors.