emergentai/cervix-type-ood

Model Card: DenseNet121 for Cervix type Image Classification

This model classifies cervical images into Type_1, Type_2, Type_3, and an Out-of-Distribution (OOD) category. It uses a DenseNet121 backbone pretrained on ImageNet and fine-tuned on cervical images, including OOD examples from Caltech101.

Model Details

• Base model: torchvision.models.densenet121 pretrained on ImageNet
• Input: RGB images (224x224)
• Output: 4 classes: ['Type_1', 'Type_2', 'Type_3', 'OOD']
• License: MIT
• Training dataset sources:
  • Cervical images: Intel MobileODT competition dataset
  • OOD images: Caltech101 dataset
• Preprocessing & Augmentation:
  • Resize to 224x224
  • Normalization (ImageNet mean & std)
  • Data augmentation: Random rotation, color jitter (brightness/contrast)

Dataset Distribution

Split	Type_1	Type_2	Type_3	OOD	Total
Train	557	532	547	424	2060
Validation	151	161	154	122	588
Test	73	88	80	54	295

Training Details

• Optimizer: Adam
• Loss: CrossEntropyLoss
• Batch size: 8
• Learning rate: 1e-5
• Epochs: 30
• Device: GPU (Tesla T4, 14GB)

Evaluation

Evaluation Metrics

Class	Precision	Recall	F1-score	Sensitivity	Specificity
OOD	1.00	1.00	1.00	1.0000	1.0000
Type_1	0.74	0.93	0.82	0.9333	0.9074
Type_2	0.85	0.51	0.64	0.5114	0.9574
Type_3	0.73	0.92	0.81	0.9189	0.8762

Overall accuracy: 0.81

Confusion Matrix

       Predicted
        OOD  T1  T2  T3
Actual
OOD      54   0   0   0
Type_1    0  56   3   1
Type_2    0  19  45  24
Type_3    0   1   5  68

Classification Report

              precision    recall  f1-score   support
OOD           1.00      1.00      1.00        54
Type_1        0.74      0.93      0.82        60
Type_2        0.85      0.51      0.64        88
Type_3        0.73      0.92      0.81        74

accuracy                           0.81       276
macro avg       0.83      0.84      0.82       276
weighted avg    0.82      0.81      0.80       276

---

How to Get Started

import torch
from torchvision import transforms, models
from PIL import Image

# Load model
model = models.densenet121(pretrained=False)
model.classifier = torch.nn.Linear(model.classifier.in_features, 4)
model.load_state_dict(torch.load("Dense_net_121.pth", map_location="cpu"))
model.eval()

# Transform
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225])
])

# Load image
image = Image.open("example.jpg").convert("RGB")
image = transform(image).unsqueeze(0)

# Predict
outputs = model(image)
probabilities = torch.softmax(outputs, dim=1)
predicted_class = torch.argmax(probabilities, dim=1).item()
confidence = probabilities[0, predicted_class].item()

class_names = ["Type_1", "Type_2", "Type_3", "OOD"]
print(f"Predicted class: {class_names[predicted_class]}, confidence: {confidence:.2f}")

````

---
## Technical Specifications

### Model Architecture

* **Backbone:** DenseNet121 pretrained on ImageNet
* **Output Layer:** Fully connected layer with 4 outputs (`Type_1`, `Type_2`, `Type_3`, `OOD`)
* **Activation:** Softmax for multi-class classification
* **Training Framework:** PyTorch
* **Loss Function:** CrossEntropyLoss
* **Data Handling:** Includes OOD images from Caltech101 along with in-distribution cervical images
* **Preprocessing & Augmentation:** Resize to 224x224, normalization (ImageNet mean/std), random rotation, color jitter

### Compute Infrastructure

* **Hardware:** Tesla T4 GPU (14GB)
* **Software:** PyTorch, torchvision, CUDA

---