tarirozw/gatekeeper_cervix_detector

Gatekeeper Cervix Detector

This model serves as a binary gatekeeper for cervical cancer screening systems. It quickly determines whether an input image is a valid cervical image before passing it to a downstream diagnostic model (MedSigLip).

Model Details

• Model Name: Gatekeeper Cervix Detector
• Base Architecture: MobileNetV3-Small
• Task: Binary classification (cervix vs not-cervix)
• Input Size: 224×224×3 (RGB)
• Input Normalization: [-1, 1] (mean = [0.5, 0.5, 0.5], std = [0.5, 0.5, 0.5])
• Output: Probability that the image contains a cervix (sigmoid output)
• Threshold: 0.70 (images below this confidence are rejected)
• License: CC BY 4.0
• Repository: [Link to this repo]

Intended Use

This model is designed to be used as the first stage in a two-stage pipeline:

1. Gatekeeper (this model): Filters out non-cervix images (random photos, other medical images, poor quality, etc.).
2. MedSigLip (downstream model): Only runs if the gatekeeper accepts the image. It performs cancer stage classification or similarity scoring.

Use Case: Maternal health screening in low-resource settings where non-experts may capture images using blind sweeps.

Performance

Test Set Results (threshold = 0.70)

• Accuracy: 99.94%
• Precision: 100.00%
• Recall/Sensitivity: 99.88%
• Specificity: 100.00%
• AUC: 1.0000
• Rejection Rate: ~52.0%

The model shows excellent generalization and very strong rejection of non-cervix images while maintaining high sensitivity on true cervix images.

Class Distribution

• Cervix: 48.0%
• Not-cervix: 52.0%
• The split is stratified across train/validation/test sets.

How to Use

Loading the Model

from transformers import AutoModel
model = AutoModel.from_pretrained("your-username/gatekeeper-cervix-detector")