An implementation of a framework for automatic adaption medical image detection which based on segmentation data.

Major Features:

• Automated analysis of the dataset, automatic adaptation to new medical segmentation datasets without user intervention.
• Automatically designs and executes a network training pipeline.
• Implementations of prevalent object detectors: e.g. 2D and 3D U-Net, Mask R-CNN, Retina Net, Retina U-Net.
• Modular and light-weight structure for backbone architecture: e.g. resnet, densenet, inceptionResNetV2.
• Dynamic patching and tiling (for training and inference) or full-size images.
• Weighted consolidation of box predictions across patch-overlaps, ensembles.
• Pipeline on data analysis, preprocessing, training, postprocessingg, evaluation, inference and visualization.
• Automatically chooses the best single model or ensemble to be used for test set prediction.

1. [Data analysis]: DatasetAnalyzer, Planner, Planner2D,
2. [Datasets]: DataLoaderBase, AbstractAugmentation, BatchGenerator2D, BatchGenerator3D
3. [Preprocessing]: GenericPreprocessor, PreprocessorFor2D
4. [Training]: GenericTrainer, Trainer, CascadeTrainer
5. [Inference]: Predictor .
6. [Evaluation]: Evaluator .
7. [Models]:models/* .
8. [Utilities]: utils/* .
9. [Bin]: bins/* .

1. Clone this repository
2. Setup package in virtual environment

   cd nodule_detector
   virtualenv -p python3 venv
   source venv/bin/activate
   pip3 install -e .
   

3. Install dependencies

   pip3 install -r requirements.txt