efecelik/medgemma-abdominal-ct-lora

MedGemma Abdominal CT LoRA

Abdominal organ classification adapter fine-tuned on OrganAMNIST (MedMNIST) using MedGemma 4B.

Identifies the primary organ or anatomical structure visible in abdominal CT axial slices across 11 classes.

Model Details

Property	Value
Base Model	google/medgemma-4b-it
Method	LoRA (Low-Rank Adaptation)
Task	Multi-class organ classification (11 classes)
Modality	Abdominal CT (axial 2D slices)
Framework	PyTorch + HuggingFace Transformers + PEFT

Training Dataset

OrganAMNIST from the MedMNIST v2 benchmark — standardized 2D axial CT slices for organ classification.

Reference: Yang et al. 2023, Scientific Data - "MedMNIST v2: A Large-Scale Lightweight Benchmark for 2D and 3D Biomedical Image Classification"

• Original dataset: ~58,850 images
• Train samples: 10,000 (curated subset)
• Validation samples: 1,000
• Image size: 28x28 pixels (MedMNIST standard, resized by processor)

Class Distribution

ID	Organ	Description
0	Bladder	Urinary bladder in the pelvis
1	Femur (left)	Proximal left femur and femoral head
2	Femur (right)	Proximal right femur and femoral head
3	Heart	Cardiac silhouette with chambers and great vessels
4	Kidney (left)	Left kidney with cortex and medulla
5	Kidney (right)	Right kidney (slightly lower due to liver)
6	Liver	Largest solid abdominal organ, right upper quadrant
7	Lung (left)	Left hemithorax pulmonary tissue
8	Lung (right)	Right hemithorax, three lobes
9	Spleen	Left upper quadrant, posterior to stomach
10	Pancreas	Retroperitoneal organ crossing midline

Training Configuration

LoRA Parameters

Parameter	Value
Rank (r)	16
Alpha	32
Dropout	0.05
Target Modules	all-linear
Task Type	CAUSAL_LM
Trainable Params	1.38B / 5.68B (24.3%)

Hyperparameters

Parameter	Value
Epochs	1
Per-device Batch Size	1
Gradient Accumulation Steps	8 (effective batch = 8)
Learning Rate	2e-4
LR Scheduler	Linear with warmup
Warmup Ratio	0.03
Max Grad Norm	0.3
Precision	bfloat16
Gradient Checkpointing	Enabled
Seed	42

Infrastructure

Property	Value
GPU	NVIDIA L4 (24 GB VRAM)
Cloud Platform	Modal serverless GPU
Training Time	~45-60 minutes

Prompt Format

Input:

Identify the primary organ or structure visible in this abdominal CT slice.

Output:

This abdominal CT slice primarily shows the Liver.

Liver (largest solid organ in the abdomen, occupying the right upper quadrant with homogeneous parenchymal density).

Usage

from transformers import AutoProcessor, AutoModelForImageTextToText
from peft import PeftModel
from PIL import Image

base_model_id = "google/medgemma-4b-it"
adapter_id = "efecelik/medgemma-abdominal-ct-lora"

processor = AutoProcessor.from_pretrained(base_model_id)
model = AutoModelForImageTextToText.from_pretrained(
    base_model_id, torch_dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_id)

image = Image.open("abdominal_ct.jpg").convert("RGB")
messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "Identify the primary organ or structure visible in this abdominal CT slice."}
    ]}
]

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

output = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(output[0], skip_special_tokens=True))

Intended Use

This adapter is part of the MedVision AI platform built for the MedGemma Impact Challenge. It is designed for:

• Medical education: Helping students learn abdominal CT anatomy and organ identification
• Clinical decision support: Assisting radiologists with organ localization
• Research: Exploring fine-tuned medical VLMs for abdominal imaging

Limitations

• Not for clinical diagnosis. This model is for educational and research purposes only.
• Organ identification only: Classifies visible organ, does not detect pathology within organs.
• Low resolution source: MedMNIST images are 28x28 pixels, limiting fine structural detail.
• Normal anatomy only: Trained on healthy organ appearances, not pathological variants.
• Single epoch: Trained for 1 epoch; further training may improve performance.

Citation

@article{yang2023medmnist,
  title={MedMNIST v2-A large-scale lightweight benchmark for 2D and 3D biomedical image classification},
  author={Yang, Jiancheng and Shi, Rui and Wei, Donglai and Liu, Zequan and Zhao, Lin and Ke, Bilian and Pfister, Hanspeter and Ni, Bingbing},
  journal={Scientific Data},
  volume={10},
  number={1},
  pages={41},
  year={2023},
  publisher={Nature Publishing Group UK London}
}

Disclaimer

This model is for educational and research purposes only. It is NOT intended for clinical diagnosis or patient care decisions. Always consult qualified medical professionals for medical advice.