Skip to content

Prior-free deep learning for pediatric bone age assessment robust to skeletal dysplasias

License

Notifications You must be signed in to change notification settings

aimi-bonn/Deeplasia

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

70 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Deeplasia

Code style: black     License: CC BY-NC 4.0

This repository contains the official code for Deeplasia. We use a simple, prior-free deep learning approach to assert generalizability to unseen and uncommon bone shapes.

The models were trained on the RSNA Pediatric Bone Age Dataset. The dataset contains ca. 14,200 images of the hands and wrists of children, with ages ranging from 0 to 18 years.

The deep learning models consist of a convolutional backbone (efficientnet or inception-v3) and a fully connected classifier performing the regression of the bone age. Usually, the model takes the sex as additional input. However, it can also be trained to predict the sex in addition to the bone age as multi-task learning (MTL).

Usage

Installation

To install the necessary dependencies, run:

$ pip install -r requirements.txt

For GPU acceleration using CUDA, you need to install the CUDA 11.3+ versions of torch and torchvision:

torch==1.10.2+cu113
torchvision==0.11.3+cu113

Please refer to the official PyTorch installation guide for more details.

Data

The RSNA Pediatric Bone Age Dataset and the Los Angeles Digital Hand Atlas are publicly available for training and testing.

Annotation formatting

To assert compatibility with varying data sources the original annotations of the RSNA dataset are converted to a common .csv file containing the annotations from all subsets. An example of the annotations file containing the RSNA Pediatric Bone Age Dataset and the Los Angeles Digital Hand Atalas (DHA) can be found in data/annotations.csv.

The hand masks are available from zenodo.

Splits

The data/splits folder contains .csv files defining the splits of the corresponding data sets. Hereby, the patient IDs match the splits to the corresponding patients in the annotations file. So far, files for the original RSNA competition on Kaggle and the DHA test set are available.

Inference

Batched inference

For batched predictions, use the predict.py script; e.g. to test a model on the DHA with the created annotation, run:

$ python predict.py \
    --ckp_path=<path/to/checkpoint.ckpt> \
    --gpus=1 \
    --annotation_csv=data/annotation.csv \
    --split_csv=data/splits/la_dha_test.csv \
    --split_column="test" \
    --split_name="test" \
    --mask_crop_size=1.15 \
    --num_workers=8 \
    [...]

For in-depth instruction on how to test (including the final ensembling) see the example notebook nb/example_inference.ipynb.

Single image inference and deployment

See out streamlit app for easy local inference. We also provide a docker image containing Deeplasia and providing a RESTful API, see deeplasia-service for details.

Training

An exemplary training can be started with:

$ python train_model.py \
    --config=configs/cosine_annealing.yml \
    --trainer.max_epochs=100 \
    --model.backbone=efficientnet-b4 \
    --data.annotation_df=data/annotations.csv \
    --data.num_workers=8 \
    --data.train_batch_size=32 \
    --data.test_batch_size=48 \
    [...]

Logging

Per default, logs are written to run.log. To specify a different path, run the script with the $LOG_FILE environment variable:

$ LOG_FILE=<path/to/log_file.txt> python train_model.py [...]

Flags and options

For all options, check the lib/datasets.py and lib/models.py files or type

$ python train_model.py -h

to obtain an overview of all flags and options. Note that for bundling jointly used options (e.g. for training on a certain device) it is possible to create a special config and add it as --config=/path/to/config.yml flag.

For general training options (i.e. everything except the model and data options) check the pytorch-lightning documentation.

Settings sex as input or output

The following configurations are available:

  • Use the sex as input (configs/sex_input.yml, as the default configuration)
  • Predict the sex explicitly (i.e. in a separate classifier) in an MTL setting (configs/explicit.yml). Hereby
    • either the predicted sex can be used for the age prediction (--model.correct_predicted_sex=False)
    • or the ground truth sex can be used for age prediction during training (--model.correct_predicted_sex=True, default)
  • Aim to predict the sex implicitly (i.e. without a separate classifier) in an MTL setting (configs/implicit.yml)
  • Predict only the sex and ignore the bone age (configs/sex_only.yml)

Training via SLURM

A template for a SLURM training job is available at bash/slurm_job_example.sh.

Citation

@article{rassmann2023deeplasia,
  title={Deeplasia: deep learning for bone age assessment validated on skeletal dysplasias},
  author={Rassmann, Sebastian and Keller, Alexandra and Skaf, Kyra and Hustinx, Alexander and Gausche, Ruth and Ibarra-Arrelano, Miguel A and Hsieh, Tzung-Chien and Madajieu, Yolande ED and N{\"o}then, Markus M and Pf{\"a}ffle, Roland and others},
  journal={Pediatric Radiology},
  pages={1--14},
  year={2023},
  publisher={Springer}
}

About

Prior-free deep learning for pediatric bone age assessment robust to skeletal dysplasias

Resources

License

Stars

Watchers

Forks