Feasibility of deep learning-based cancer detection in ultrasound microvascular images: Dataset

This repository provides the data and results for an upcoming publication that evaluates the feasibility of applying convolutional neural networks towards cancer detection in acoustic angiography volumes, acquired in vivo. The associated code for model training and hyperparameter tuning in a nested k-fold cross-validation study can be found in the linked GitHub repository. 

The all_data.zip file contains three primary folders outlined below:

• FOR_TRAINING - .npz files containing 2-D or 3-D labeled datasets, packaged for training
  • 2d_data.npz: lateral-elevation maximum intensity projections
  • 3d_data.npz: full volumes
• MAT_FILES - raw .mat files for individual volume acquisitions that can be loaded into MATLAB 
  
  • bmodes - high frequency B-modes of the tumor volumes
    • each volume acquisition .mat file is prefixed with "PR" containing:
      • ax: axial dimension array [mm]
      • lat: lateral dimension array [mm]
      • scan: elevation dimension array [mm]
      • bmode: bmode volume, axial x lateral x elevation
      • imtor: acoustic angiography volume, axial x lateral x elevation
      • tormip: lateral-elevation maximum intensity projection
    • files2dataset.mat associates the .mat files in the folder to the packaged .npz datasets
      
      • files: a cell array of file names
      • dataset_idx: 0-based indexes corresponding to the data in the .npz datasets
      • tumor_sizes: measured tumor sizes [mm], defined as the longest diameter of the tumor
  • controls
    
    • each volume acquisition .mat file is prefixed with "PR" containing:
      
      • out - a structure containing the following fields:
        
        • ax: axial dimension array [mm]
        • lat: lateral dimension array [mm]
        • scan: elevation dimension array [mm]
        • imtor: acoustic angiography volume, axial x lateral x elevation
    • files2dataset.mat associates the .mat files in the folder to the packaged .npz datasets
      
      • files: a cell array of file names
      • dataset_idx: 0-based indexes corresponding to the data in the .npz datasets
  • tumors
    
    • each volume acquisition .mat file is prefixed with "PR" containing:
      
      • out - a structure containing the following fields:
        
        • ax: axial dimension array [mm]
        • lat: lateral dimension array [mm]
        • scan: elevation dimension array [mm]
        • imtor: acoustic angiography volume, axial x lateral x elevation
    • files2dataset.mat associates the .mat files in the folder to the packaged .npz datasets
      
      • files: a cell array of file names
      • dataset_idx: 0-based indexes corresponding to the data in the .npz datasets
• TRAINED_MODELS - folders containing the final trained 2-D and 3-D models of the nested k-fold cross-validation study (for each outer fold)
  
  • best_acc.config - model configuration with the best accuracy after hyperparameter tuning
  • best_acc.pt - model with the best accuracy after final training on the outer fold
  • best_loss.config - model configuration with the best loss after hyperparameter tuning
  • best_loss.pt - model with the best loss after final training on the outer fold
  • val_acc.dat - file to keep track of the best validation accuracy score
  • val_loss.dat - file to keep track of the best validation loss score