subvisible

Subvisible Approach

A parameter efficient solution for automatic growth plate detection in micro-CT mice bone scans

The solution is based on the observation that prior to reaching the Growth Plate Plane (GPP), a series of images in the axial plane present four distinct blobs, each corresponding to a protrusion. These blobs begin to merge upon reaching the GPP. We trained a binary classifier to identify images with these four blobs. This classifier identifies a series of images preceding the GPP as positive cases i.e. containing four blobs. The last index in this series serves as a rough estimate of the GPP.

To refine these initial estimates, we leveraged information on the axial axis and performed a search around these points. For this purpose, we employed a regression CNN. We examined 25 slices before and after the initial rough estimates, resulting in a stack of 51 images. This stack was fed into the model, the aim of which was to generate refined predictions of the GPP.

Figure 1 shows the overall diagram of our proposed approach. Considering the limited amount of training data, we trained two lightweight CNN networks for both the classification and regression tasks. Each network as shown in Figure 2 has only four levels of convolution layers and less than 40K parameters in total. These lightweight networks, along with resizing the images to 96×96 before feeding them to the models, make the training process quite fast. The loss function for the classification network is cross-entropy loss, and the negative of the score function is used as the loss function for the regression network.

Figure 1

Figure 2

Installation

Make sure you have the required dependencies installed: pip install -r requirements.txt

Training the classifier

from models import *
classifier = ImageClassifier()
classifier.train_cls(trn_folder, val_folder, logdir, checkpoint_filepath)
"""
trn_folder: Path to the training folder. This folder should contain two subfolders: 
  - "1": For positive images
  - "0": For negative images
val_folder: Path to the validation folder. This folder should also contain two subfolders:
  - "1": For positive images
  - "0": For negative images
logdir: Path to a directory for saving logs.
checkpoint_filepath: Path to a directory for saving the trained model.
"""

Training the regressors

It is assumed that the image paths and their labeled growth plate indices, along with their assigned folds for training, are saved in a CSV file. Please ensure the column names are "Growth Plate Index", "image_path", and "fold".

from models import *
data=pd.read_csv('path/to/csv.csv')
train(data, cls_checkpoint, log_path)
"""
cls_checkpoint: Path to the trained model for the classification (from the previous part) 
log_path: Path to a directory for saving logs and checkpoints. Logs get saved in the 'logs'
          subfolder and checkpoints in the 'checkpoints' subfolder. There will be separate
          folders for each fold 
"""

Inference

from models import *
result=make_prediction(path, cls_checkpoint_filepath, reg_checkpoints_dir)
"""
parameters:
  path: Path to a directory containing .nii images
  cls_checkpoint_filepath: Path to the trained model for the classification
  reg_checkpoints_dir: Path to the trained regression models, this will be
                       log_path (from the previous part) + 'checkpoints/'
returns:
a dataframe containing image names and their predicted growth plate index
"""