vit-base-patch16-dinov3-finetuned-skin-lesion-classification

vit-base-patch16-dinov3-finetuned-skin-lesion-classification

This model is a finetuned for skin lesion classification.

Params (M): 85.6

Intended Uses & Limitations

Intended Use

This model is intended for dermoscopic skin lesion classification using a 224x224 image size.

Limitations

This model was only trained for 1 epoch and has not seen many malignant examples (due to large class imbalance in ISIC 2024 dataset).

How to Get Started with the Model

class DinoSkinLesionClassifier(nn.Module, PyTorchModelHubMixin):
  """
  PytorchModelHubMixin adds push to Hugging Face Hub

  See: https://huggingface.co/docs/hub/models-uploading#upload-a-pytorch-model-using-huggingfacehub
  """
  def __init__(self, num_classes=1, freeze_backbone=True):
    super(DinoSkinLesionClassifier, self).__init__()

    # Initialize Dino v3 backbone
    self.backbone = timm.create_model('vit_base_patch16_dinov3', pretrained=True, num_classes=0, global_pool='avg')

    # Freeze backbone weights if requested
    # This makes training much faster
    if freeze_backbone:
      for param in self.backbone.parameters():
        param.requires_grad = False

    # Get feature dimension from the backbone
    feat_dim = self.backbone.num_features

    # Define the classification head
    self.head = nn.Linear(feat_dim, num_classes) # Should be 768 in, 1 out

  def forward(self, x):
    out = self.backbone(x)
    out = self.head(out)

    return out

from huggingface_hub import hf_hub_download

weights_path = hf_hub_download(
        repo_id="avanishd/vit-base-patch16-dinov3-finetuned-skin-lesion-classification",
        filename="model.safetensors"
    )

from safetensors.torch import load_model

model = EfficientNetSkinLesionClassifier()
load_model(model, filename=weights_path, strict=True)

model.to(device) # Don't forget to put on GPU

model.eval()  # Set model to evaluation mode


# Example with PH2 Dataset

class PH2Dataset(Dataset):
  """
  Dataset for PH2 images, which are in png format.

  PH2 contains skin lesions images classified as

  - Common Nevus (benign)
  - Atypical Nevus (benign)
  - Melanoma (malignant)

  No need for is real label here, since this is purely for testing
  """

  def __init__(self, dir_path, metadata, transform=None):
    super(PH2Dataset, self).__init__()

    self.dir_path = dir_path
    self.transform = transform

    self.image_files = [os.path.join(dir_path, f) for f in os.listdir(dir_path)
                        if f.lower().endswith(('.jpg', '.jpeg', '.png'))]

    # Load metadata w/ polars (only 2 columns)
    self.metadata = pl.read_csv(metadata)

    self.diagnostic_mapping = {
        "Common Nevus": 0,
        "Atypical Nevus": 0,
        "Melanoma": 1,
    }

  def __len__(self):
    return len(self.image_files)

  def __getitem__(self, idx):
    # The image name in the metadata csv are like IMD003
    image_id = self.image_files[idx].split('/')[-1].split('.')[0]

    # Still need the entire path to open the image
    image = Image.open(self.image_files[idx]).convert('RGB')

    if self.transform: # Apply transform if it exists
      image = self.transform(image)

    diagnosis = self.metadata.filter(pl.col("image_name") == image_id).select("diagnosis").item()

    label = torch.tensor(self.diagnostic_mapping[diagnosis], dtype=torch.int16)

    return image, label


transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]), # Image net mean and std
    transforms.Resize((224, 224)),  # Dimensions for Efficient Net v2
])

ph_2_images = "/content/data/ph2_data/images"
ph_2_metadata = "/content/data/ph2_data/ph_2_dataset.csv"

ex_dataset = PH2Dataset(ph_2_images, ph_2_metadata, transform)

ex_loader = DataLoader(ex_dataset, batch_size=64, shuffle=False)

for (images, labels) in test_loader:
  images = images.to(device)
  labels = labels.to(device)
  output = model(images)

  y_pred_prob = torch.sigmoid(output).cpu().numpy().ravel()
  y_pred = np.where(y_pred_prob < 0.5, 0, 1)

  return y_pred

Training and evaluation data

This model was trained with the ISIC 2024 challenge and ISIC 2024 synthetic datasets.

For the ISIC 2024 Challenge data, an 80-20 train test split was applied, and the test split was used to evaluate the model.

Training Procedure

Training hyperparameters

• learning_rate: 1e-4
• train_batch_size: 64
• eval_batch_size: 64
• optimizer: Use OptimizerNames.ADAMW_TORCH with weight decay=1e-2 and optimizer_args=No additional optimizer arguments
• num_epochs: 1

Training results

Framework versions

• Pytorch 2.9.0+cu126
• torchvision: 0.24.0+cu126
• timm: 1.0.22
• numpy: 2.0.2
• safetensors: 0.7.0