model_card.json

{
  "model_name": "PRIMER",
  "full_name": "Pretrained RadImageNet for Mammography Embedding Representations",
  "version": "1.0.0",
  "release_date": "2024-10-17",

  "model_details": {
    "organization": "Lab-Rasool",
    "architecture": "ResNet-50",
    "base_model": "RadImageNet ResNet-50",
    "training_method": "SimCLR Contrastive Learning",
    "model_type": "Feature Extraction / Embedding Model",
    "modality": "Medical Imaging - Mammography",
    "parameters": "23.5M",
    "model_size_mb": 283,
    "license": "Apache-2.0"
  },

  "intended_use": {
    "primary_uses": [
      "Feature extraction for mammography images",
      "Similarity search and retrieval",
      "Clustering and grouping mammograms",
      "Transfer learning backbone for downstream tasks",
      "Content-based image retrieval systems",
      "Quality control and anomaly detection"
    ],
    "out_of_scope": [
      "Direct clinical diagnosis",
      "Standalone diagnostic tool",
      "Non-mammography medical images",
      "Real-time processing without optimization"
    ]
  },

  "training_data": {
    "dataset": "CMMD (Chinese Mammography Mass Database)",
    "dataset_url": "https://doi.org/10.7937/tcia.eqde-3b16",
    "num_training_samples": 13000,
    "data_splits": {
      "train": 0.7,
      "validation": 0.15,
      "test": 0.15
    },
    "image_format": "DICOM",
    "views": ["CC (craniocaudal)", "MLO (mediolateral oblique)"],
    "population": "Chinese population"
  },

  "training_procedure": {
    "method": "Self-supervised contrastive learning (SimCLR)",
    "loss_function": "NT-Xent (Normalized Temperature-scaled Cross Entropy)",
    "epochs": 50,
    "batch_size": 128,
    "optimizer": "AdamW",
    "learning_rate": 0.0001,
    "scheduler": "Cosine annealing with warmup",
    "temperature": 0.07,
    "mixed_precision": true,
    "hardware": "NVIDIA RTX 3090 (24GB VRAM)"
  },

  "performance_metrics": {
    "embedding_quality": {
      "silhouette_score": {
        "radimagenet_baseline": 0.127,
        "primer_finetuned": 0.289,
        "improvement_percent": 127
      },
      "davies_bouldin_score": {
        "radimagenet_baseline": 2.847,
        "primer_finetuned": 1.653,
        "improvement_percent": -42,
        "note": "Lower is better"
      },
      "calinski_harabasz_score": {
        "radimagenet_baseline": 1834,
        "primer_finetuned": 3621,
        "improvement_percent": 97
      },
      "embedding_variance": {
        "radimagenet_baseline": 0.012,
        "primer_finetuned": 0.024,
        "improvement_percent": 100
      }
    }
  },

  "input_output": {
    "input": {
      "format": "DICOM or preprocessed image tensor",
      "shape": [3, 224, 224],
      "dtype": "float32",
      "color_space": "RGB",
      "normalization": "ImageNet (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])"
    },
    "output": {
      "format": "Embedding vector",
      "shape": [2048],
      "dtype": "float32",
      "normalization": "L2 normalization recommended"
    }
  },

  "preprocessing_requirements": {
    "critical_steps": [
      "Photometric interpretation correction (MONOCHROME1 inversion)",
      "Percentile-based intensity clipping (2nd-98th percentile)",
      "Min-max normalization to [0, 255]",
      "CLAHE enhancement (clipLimit=2.0, tileGridSize=8x8)",
      "Grayscale to RGB conversion",
      "Resize to 224x224",
      "ImageNet normalization"
    ],
    "dependencies": [
      "pydicom>=2.4.4",
      "opencv-python>=4.8.1.78",
      "numpy>=1.26.0"
    ]
  },

  "limitations": {
    "domain_specificity": "Trained on CMMD dataset (Chinese population); performance may vary on other populations",
    "dicom_dependency": "Requires proper DICOM preprocessing for optimal results",
    "resolution_loss": "High-resolution details may be lost at 224x224 input size",
    "self_supervised": "No direct classification output; requires downstream task integration",
    "photometric_interpretation": "Critical to handle MONOCHROME1/MONOCHROME2 correctly"
  },

  "ethical_considerations": {
    "bias": "Model trained on Chinese population data; may not generalize equally to all demographics",
    "clinical_use": "Not FDA approved; requires clinical validation before medical use",
    "privacy": "DICOM files may contain PHI; ensure proper de-identification",
    "interpretability": "Embeddings are learned representations; clinical interpretation required"
  },

  "citations": {
    "primer": {
      "title": "PRIMER: Pretrained RadImageNet for Mammography Embedding Representations",
      "authors": "Lab-Rasool",
      "year": 2024,
      "url": "https://huggingface.co/Lab-Rasool/PRIMER"
    },
    "radimagenet": {
      "title": "RadImageNet: An Open Radiologic Deep Learning Research Dataset for Effective Transfer Learning",
      "authors": "Mei et al.",
      "journal": "Radiology: Artificial Intelligence",
      "year": 2022,
      "doi": "10.1148/ryai.210315"
    },
    "simclr": {
      "title": "A Simple Framework for Contrastive Learning of Visual Representations",
      "authors": "Chen et al.",
      "conference": "ICML",
      "year": 2020,
      "arxiv": "2002.05709"
    },
    "cmmd": {
      "title": "Chinese Mammography Database (CMMD)",
      "source": "The Cancer Imaging Archive",
      "doi": "10.7937/tcia.eqde-3b16"
    }
  },

  "contact": {
    "organization": "Lab-Rasool",
    "huggingface": "https://huggingface.co/Lab-Rasool",
    "model_repository": "https://huggingface.co/Lab-Rasool/PRIMER",
    "issues": "https://huggingface.co/Lab-Rasool/PRIMER/discussions"
  },

  "technical_specifications": {
    "framework": "PyTorch 2.1+",
    "required_libraries": [
      "torch>=2.1.0",
      "torchvision>=0.16.0",
      "timm>=0.9.12",
      "pydicom>=2.4.4",
      "opencv-python>=4.8.1.78",
      "albumentations>=1.3.1"
    ],
    "gpu_requirements": "12GB+ VRAM recommended for inference",
    "inference_speed": "~50ms per image on RTX 3090"
  }
}