This project provides a complete pipeline to train a machine learning model capable of classifying DICOM images. The process uses a pre-trained ResNet50 model to extract visual features from the images and a Support Vector Classifier (SVC) to perform the final classification.

The workflow is divided into four main phases, each corresponding to a Python script.

.
├── input_dicom/               # <-- Place your original, unsorted DICOM files here (can have subfolders)
│   ├── folder1/
│   │   └── original_image1
│   └── original_image2
├── datos_etiquetados/           # (Generated by you after Phase 1)
│   ├── FDT/                     # <-- Place "FDT" class images here
│   ├── MTF/                     # <-- Place "MTF" class images here
│   └── TOR/                     # <-- Place "TOR" class images here
├── dataset_final/               # (Generated by Phase 2)
│   ├── dataset_X.npy
│   └── dataset_y.npy
├── modelo_final/                # (Generated by Phase 3)
│   ├── clasificador_svc.joblib
│   ├── escalador.joblib
│   └── matriz_de_confusion.png
├── fase1_descompresion.py       # Script for Phase 1
├── fase2_generar_vectores.py    # Script for Phase 2
├── fase3_entrenamiento_svc.py   # Script for Phase 3
├── fase4_inferencia.py          # Script for Phase 4 (inference)
├── config.py                    # Simplified configuration file
├── utils.py                     # Utility functions
└── pyproject.toml               # Project dependencies

This project uses Python 3.13 or higher. The required libraries are listed in the pyproject.toml file. You can install them using a package manager like pip or uv.

# Example installation with pip
pip install numpy torch torchvision pillow pydicom scikit-learn seaborn joblib gdcm pylibjpeg pylibjpeg-libjpeg

Follow these phases in order to go from raw DICOM files to a fully functional classifier.

This phase reads all valid DICOM files from the input_dicom directory (including subdirectories), decompresses them, and renames them in chronological order based on their acquisition time.

➡️ Action:

1. Place all your original DICOM files (with or without extension) inside the input_dicom folder. You can organize them in subfolders if you wish.
2. Run the script from your terminal:

   python fase1_descompresion.py

➡️ Result:

• A new folder named f1_descomprimidos will be created, containing the standardized DICOM files. The files will be named chronologically (e.g., Img1_..., Img2_...).

This is a manual step where you provide the ground truth for the model.

➡️ Action:

1. Create a folder named datos_etiquetados.
2. Inside datos_etiquetados, create subfolders for each of your classes (e.g., FDT, MTF, TOR).
3. Move the files from f1_descomprimidos into their corresponding class folder inside datos_etiquetados.

This script processes the labeled images, applies a powerful data augmentation strategy, and uses a pre-trained ResNet50 model to extract a feature vector for each image. At the end, it consolidates all data into two master files for training.

The script uses a differentiated data augmentation strategy:

• Aggressive Augmentation for classes with objects (MTF, TOR), including full rotations and large translations, to teach the model position/rotation invariance.
• Gentle Augmentation for the background class (FDT), focusing on brightness and contrast changes to preserve the background structure.

➡️ Action:

1. Ensure you have completed Phase 1.5 and your labeled images are in the datos_etiquetados folder.
2. Run the script:

   python fase2_generar_vectores.py

➡️ Result:

• A new folder named dataset_final will be created, containing:
  • dataset_X.npy: An array with all the feature vectors.
  • dataset_y.npy: An array with the corresponding label for each vector.
• An output_procesado_aumentado folder is also created for verification purposes.

This script takes the consolidated dataset from Phase 2, splits it into training and testing sets, and trains a Support Vector Classifier (SVC) model.

➡️ Action:

1. Ensure the dataset_final folder with its .npy files exists.
2. Run the training script:

   python fase3_entrenamiento_svc.py

➡️ Result:

• The script will print a Classification Report and Global Accuracy to the console, showing the model's performance on the test data.
• A new folder modelo_final will be created, containing:
  • clasificador_svc.joblib: The trained and saved AI model.
  • escalador.joblib: The saved data scaler, crucial for inference.
  • matriz_de_confusion.png: A visual plot of the model's performance.

This final script uses the trained model to classify a new, unseen DICOM image.

➡️ Action:

1. Place a new DICOM image you want to classify in the project's root directory (or any other location).
2. Open the fase4_inferencia.py script and modify the RUTA_IMAGEN_NUEVA variable to point to your new image file.
3. Run the inference script:

   python fase4_inferencia.py

➡️ Result:

• The script will print the final prediction to the console, showing the classified label and the confidence score for each class. For example:

  ✅ La imagen ha sido clasificada como: 'FDT'
  
  📊 Confianza por clase:
    - FDT: 98.14%
    - MTF: 1.55%
    - TOR: 0.31%