DOSMA: Deep Open-Source Medical Image Analysis

Documentation | Questionnaire | DOSMA Basics Tutorial

DOSMA is an AI-powered Python library for medical image analysis. This includes, but is not limited to:

image processing (denoising, super-resolution, registration, segmentation, etc.)
quantitative fitting and image analysis
anatomical visualization and analysis (patellar tilt, femoral cartilage thickness, etc.)

We hope that this open-source pipeline will be useful for quick anatomy/pathology analysis and will serve as a hub for adding support for analyzing different anatomies and scan sequences.

Installation

DOSMA requires Python 3.6+. The core module depends on numpy, nibabel, nipype, pandas, pydicom, scikit-image, scipy, PyYAML, and tqdm.

Additional AI features can be unlocked by installing tensorflow and keras. To enable built-in registration functionality, download elastix. Details can be found in the setup documentation.

To install DOSMA, run:

pip install dosma

# To install with AI support
pip install dosma[ai]

If you would like to contribute to DOSMA, we recommend you clone the repository and install DOSMA with pip in editable mode.

git clone git@github.com:ad12/DOSMA.git
cd DOSMA
pip install -e '.[dev,docs]'
make dev

To run tests, build documentation and contribute, run

make autoformat test build-docs

Features

Simplified, Efficient I/O

DOSMA provides efficient readers for DICOM and NIfTI formats built on nibabel and pydicom. Multi-slice DICOM data can be loaded in parallel with multiple workers and structured into the appropriate 3D volume(s). For example, multi-echo and dynamic contrast-enhanced (DCE) MRI scans have multiple volumes acquired at different echo times and trigger times, respectively. These can be loaded into multiple volumes with ease:

import dosma as dm

multi_echo_scan = dm.load("/path/to/multi-echo/scan", group_by="EchoNumbers", num_workers=8, verbose=True)
dce_scan = dm.load("/path/to/dce/scan", group_by="TriggerTime")

Data-Embedded Medical Images

DOSMA's MedicalVolume data structure supports array-like operations (arithmetic, slicing, etc.) on medical images while preserving spatial attributes and accompanying metadata. This structure supports NumPy interoperability, intelligent reformatting, fast low-level computations, and native GPU support. For example, given MedicalVolumes mvA and mvB we can do the following:

# Reformat image into Superior->Inferior, Anterior->Posterior, Left->Right directions.
mvA = mvA.reformat(("SI", "AP", "LR"))

# Get and set metadata
study_description = mvA.get_metadata("StudyDescription")
mvA.set_metadata("StudyDescription", "A sample study")

# Perform NumPy operations like you would on image data.
rss = np.sqrt(mvA**2 + mvB**2)

# Move to GPU 0 for CuPy operations
mv_gpu = mvA.to(dosma.Device(0))

# Take slices. Metadata will be sliced appropriately.
mv_subvolume = mvA[10:20, 10:20, 4:6]

Built-in AI Models

DOSMA is built to be a hub for machine/deep learning models. A complete list of models and corresponding publications can be found here. We can use one of the knee segmentation models to segment a MedicalVolume mv and model weights downloaded locally:

from dosma.models import IWOAIOAIUnet2DNormalized

# Reformat such that sagittal plane is last dimension.
mv = mv.reformat(("SI", "AP", "LR"))

# Do segmentation
model = IWOAIOAIUnet2DNormalized(input_shape=mv.shape[:2] + (1,), weights_path=weights)
masks = model.generate_mask(mv)

Parallelizable Operations

DOSMA supports parallelization for compute-heavy operations, like curve fitting and image registration. Image registration is supported thru the elastix/transformix libraries. For example we can use multiple workers to register volumes to a target, and use the registered outputs for per-voxel monoexponential fitting:

# Register images mvA, mvB, mvC to target image mv_tgt in parallel
_, (mvA_reg, mvB_reg, mvC_reg) = dosma.register(
   mv_tgt,
   moving=[mvA, mvB, mvC],
   parameters="/path/to/elastix/registration/file",
   num_workers=3,
   return_volumes=True,
   show_pbar=True,
)

# Perform monoexponential fitting.
def monoexponential(x, a, b):
   return a * np.exp(b*x)

fitter = dosma.CurveFitter(
   monoexponential,
   num_workers=4,
   p0={"a": 1.0, "b": -1/30},
)
popt, r2 = fitter.fit(x=[1, 2, 3, 4], [mv_tgt, mvA_reg, mvB_reg, mvC_reg])
a_fit, b_fit = popt[..., 0], popt[..., 1]

Citation

@inproceedings{desai2019dosma,
  title={DOSMA: A deep-learning, open-source framework for musculoskeletal MRI analysis},
  author={Desai, Arjun D and Barbieri, Marco and Mazzoli, Valentina and Rubin, Elka and Black, Marianne S and Watkins, Lauren E and Gold, Garry E and Hargreaves, Brian A and Chaudhari, Akshay S},
  booktitle={Proc 27th Annual Meeting ISMRM, Montreal},
  pages={1135},
  year={2019}
}

In addition to DOSMA, please also consider citing the work that introduced the method used for analysis.

ad12 / DOSMA