Anidane (MedGemma 1.5 4B Fine-Tuned for Ophthalmic Analysis)

Anidane is a fine-tuned, multimodal vision-language model optimized for the analysis and diagnostic assessment of ophthalmic imagery, specifically fundus and retinal photography. Built on top of the google/medgemma-1.5-4b-it architecture and accelerated using Unsloth, Anidane bridges the gap in clinical accessibility by enabling parameter-efficient interpretation of localized eye pathologies and structural abnormalities.

Model Details

• Developed by: WILLIAMS OPOKU & SAMUEL AGYEMANG
• Model Type: Multimodal Vision-Language Model (Image-Text to Text)
• Base Model: google/medgemma-1.5-4b-it (Gemma 3 / MedSigLIP Core)
• Fine-Tuning Method: QLoRA (4-bit quantized) via Unsloth
• Primary Domain: Ophthalmology, Retinal Image Diagnostic Reporting

Intended Uses & Limitations

Intended Use

• Clinical Assistant Tool: Providing detailed descriptive impressions of fundus abnormalities (e.g., microaneurysms, vascular alterations, optic disc changes).
• Research & Education: Analyzing pathological trends in public ophthalmic imagery datasets, with an emphasis on mitigating geographic screening biases in underrepresented cohorts.

Limitations & Out of Scope

• Not a Standalone Diagnostic Tool: Anidane is designed strictly as an assistive reporting tool and must not be used for definitive medical diagnoses without certified practitioner oversight.
• Data Demographics: While optimized to interpret global clinical profiles, the model's accuracy is closely tied to the imaging quality and distribution found within the underlying train sets.

Training Details

Training Dataset

• Dataset Source: smrndmdude/eye-diseases-dataset
• Modality: High-resolution eye/fundus photography paired with localized clinical diagnostic captions.

Hyperparameters & Optimization

The model was fine-tuned using the following configurations:

• Quantization: 4-bit NormalFloat4 (NF4)
• LoRA Rank ($r$): 16
• LoRA Alpha ($\alpha$): 16
• Optimizer: adamw_8bit
• Learning Rate: $2 \times 10^{-4}$ (Linear decay)
• Gradient Accumulation Steps: 4
• Batch Size: 2 per device
• Total Steps: 59

Training Loss Curve Performance

During fine-tuning, the model displayed highly stable convergence characteristics, with training loss scaling down seamlessly as follows:

• Initial Phase (Step 1): 6.061622
• Mid Phase (Step 30): 0.171660
• Final Convergence Phase (Step 59): 0.080335

How to Use the Model

You can load and query Anidane using the Hugging Face transformers library or via unsloth for accelerated inference.

import torch

from PIL import Image

from transformers import AutoProcessor

from unsloth import FastVisionModel # Import FastVisionModel



model_id = "Willie999/Anidane"



processor = AutoProcessor.from_pretrained(model_id)

# Use FastVisionModel for loading the model, consistent with training

# FastVisionModel.from_pretrained returns a tuple (model, processor), so we unpack it.

model, _ = FastVisionModel.from_pretrained(

    model_name = model_id, # Use model_name for FastVisionModel

    torch_dtype=torch.bfloat16,

    device_map="auto"

)



# Load a target retinal image

image = Image.open("/content/red-eye-disease.jpg").convert("RGB")



messages = [

    {

        "role": "user",

        "content": [

            {"type": "image"},

            {"type": "text", "text": "Analyze this fundus/eye photograph. Provide a clear diagnostic assessment focusing on localized anomalies."}

        ]

    }

]



prompt = processor.apply_chat_template(messages, add_generation_prompt=True)

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



with torch.no_grad():

    generated_ids = model.generate(

        input_ids=inputs["input_ids"],

        pixel_values=inputs["pixel_values"], # Explicitly pass pixel_values

        max_new_tokens=256,

        do_sample=False

    )



response = processor.batch_decode(

    generated_ids[:, inputs["input_ids"].shape[1]:],

    skip_special_tokens=True

)[0]



print("Anidane Analysis:", response)