Towards subject-level cerebral infarction classification of CT scans using convolutional networks

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Neurology

Background:

• Automatic evaluation of 3D medical volumes is crucial for efficient clinical decision-making.
• Classifying non-acute cerebral infarction in computed tomography (CT) scans is challenging due to lesion similarity in shape and intensity.
• Existing methods may lack transparency, hindering clinical trust and assessment.

Purpose of the Study:

• To develop and validate a robust, multi-stage automated method for classifying non-acute cerebral infarction in full 3D CT volumes.
• To improve the speed and accuracy of diagnosing cerebral infarction.
• To provide a transparent classification approach with assessable intermediate results.

Main Methods:

• A three-stage architecture was employed: cranial cavity segmentation, region proposal generation, and multi-resolution, densely connected 3D convolutional network classification.
• The method was trained and tested on a dataset of 555 CT scans, with 186 used for the unstratified test set.
• Dataset bias related to patient age and CT scanner model was investigated.

Main Results:

• The proposed method achieved high subject-level classification performance, with a mean area under the curve (AUC) of 0.95 for the unstratified test set.
• Performance remained strong when stratified by patient age (AUC 0.88) and CT scanner model (AUC 0.93).
• The system demonstrated successful full volume classification and offered transparency through intermediate segmentation results.

Conclusions:

• The developed three-stage model effectively classifies non-acute cerebral infarction in CT volumes automatically.
• The approach offers a transparent alternative to black-box methods, allowing for examination of intermediate segmentation outputs.
• This automated method has the potential to significantly expedite clinical decision-making in neuroimaging.