Anatomy-Specific Classification of Medical Imaging Data Hosted by Medical Imaging and Data Resource Center (MIDRC)
N Shusharina1*, D Krishnaswamy2, P Kinahan3, A Fedorov2, (1) Massachusetts General Hospital, Boston, MA, (2) Brigham & Women's Hospital, Boston, MA, (3) University of Washington, Seattle, WA
Presentations
PO-GePV-M-25 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Medical Imaging and Data Resource Center (MIDRC) aims to build an artificial intelligence-ready data commons to support COVID imaging research. Unfortunately, the data collected by the repository often lacks critical metadata needed to build cohorts and facilitate data discovery. In particular, assignment of the image anatomy is often removed in the process of image de-identification, or is not sufficiently precise. To remedy this, we present a preliminary evaluation of a tool for automated anatomy classification on the MIDRC data.
Methods: The tool is being evaluated using CT imaging data available from the MIDRC, which was loaded to the Google Cloud Platform, with the experiments performed using Google Colab. We adopted the existing open-source “BodyPartRegression” deep-learning-based tool for automated per-slice body part detection. The output of the code was further processed to partition the entire CT volume into sub-volumes containing specific anatomical regions, including head, shoulder & neck, chest, abdomen, pelvis, and legs.
Results: Representative CT series, both whole-body and narrow focused were processed to define the image sub-volumes. The cases were selected to represent heterogeneity of the data with respect to subject gender and body weight, image resolution and slice spacing, and CT scanner model. The evaluation of the accuracy of defining the chest region was performed by automatically segmenting the lungs using image intensity thresholding and ensuring that the entirety of the lungs is included in the chest sub-volume. Similarly, the head regions were evaluated by automatically segmenting the skull. The accuracy was quantified by percentage of non-overlapping image slices.
Conclusion: We demonstrated the feasibility of automated anatomy-specific classification of MIDRC CT images. Our next steps will involve application of the tool to classify a larger data set and integration of the results within the MIDRC portal.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email