Efficient Quality Control with Mixed CT and CTA Datasets | JoVE Visualize

Area of Science:

Medical Imaging Analysis
Artificial Intelligence in Healthcare
Radiology and Neuroradiology

Background:

Deep learning models require high-quality data for analyzing large traumatic brain injury (TBI) computed tomography (CT) datasets.
Manual quality assurance (QA) of CT scans is time-consuming and often necessary to remove scans with artifacts or computed tomography angiography (CTA).
Efficient automated methods are needed to preprocess heterogeneous CT datasets for deep learning applications in TBI.

Purpose of the Study:

To develop and evaluate an automated quality assurance (QA) approach for identifying and retrieving CT scans without artifacts from TBI datasets.
To differentiate between CT and CTA scans and detect the presence of streaking artifacts using a deep learning model.
To improve the efficiency and accuracy of data preprocessing for deep learning in TBI research.

Main Methods:

A multi-headed convolutional neural network was employed to analyze 3D CT scans represented as 2D axial slice montages.
The model was trained to detect CT versus CTA and artifact versus no artifact.
A 4-fold stratified cross-validation was performed on 698 montages, with an additional 150 montages reserved for external validation.

Main Results:

The main model achieved high performance in CT detection (AUC 0.978 during cross-validation, 0.965 on external validation).
Artifact detection performance showed moderate results (AUC 0.675 during cross-validation, 0.698 on external validation).
An ablated model demonstrated slightly lower CT detection but comparable artifact detection performance, suggesting the importance of specific model components.

Conclusions:

The proposed automated QA approach is effective for CT detection in TBI datasets.
Artifact detection performance may be limited by the heterogeneity of artifacts and the number of artifact samples in the training data.
Further refinement of the model and training data is recommended to enhance artifact detection accuracy for robust deep learning applications in TBI.