| --- |
| license: mit |
| metrics: |
| - accuracy |
| - precision |
| - recall |
| - f1 |
| pipeline_tag: image-classification |
| tags: |
| - medical |
| - cervical-cancer |
| - multi-class |
| - ood-detection |
| --- |
| |
| # 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 |
|
|
| ``````python |
| 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 |
|
|
| --- |
