This model is a fine-tuned MobileViT-Small designed to classify individual cervical cells into 5 distinct morphological categories. It serves as the classification engine for the broader Cervical Cytology AI Workflow Simulation.

⚠️ Disclaimer: This model is a research proof-of-concept trained on the SIPaKMeD dataset. It is NOT a certified medical device and should not be used for clinical diagnosis.

🎯 Model Intention

This model is designed to be lightweight enough for edge deployment (simulating digital pathology integration) while maintaining high sensitivity for abnormal cell types.

📊 Class Labels & Performance

The model classifies cells in 224x224 frame of a slide into:

ID	Label	Description
0	`dyskeratotic`	Abnormal (Squamous cell carcinoma)
1	`koilocytotic`	Abnormal (HPV infection indicator)
2	`metaplastic`	Benign (Normal variation)
3	`parabasal`	Benign (Normal variation)
4	`superficial_intermediate`	Benign (Normal variation)

Metrics (Test Set):

Accuracy: ~92%
Macro F1: 93%
Recall (Abnormal Classes): 93%

🛠 Usage

python

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image

# Load Model
repo_name = "AurevinP/cervical-cytology-mobilevit-sipakmed"
processor = AutoImageProcessor.from_pretrained(repo_name)
model = AutoModelForImageClassification.from_pretrained(repo_name)

# Inference
image = Image.open("path/to/slide_patch.jpg")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
logits = outputs.logits
predicted_class = logits.argmax(-1).item()

print(f"Prediction: {model.config.id2label[predicted_class]}")

⚙️ Training Details

Dataset: SIPaKMeD (Syrto et al.)
Preprocessing: Color normalization, Center Crop, Resize.
Optimizer: AdamW / SGD
Loss: CrossEntropy with Class Weights (to handle imbalance)

📜 License

MIT License