A multi-label classification model for full slice brain computerised tomography image.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Brain computerized tomography (CT) scans are crucial for detecting brain trauma and conditions.
• Current deep learning methods for brain CT analysis often focus on individual slices, ignoring inter-slice dependencies and multi-disease complexities.
• Slice-by-slice labeling is time-consuming, expensive, and may miss concurrent conditions.

Purpose of the Study:

• To develop a novel deep learning model for accurate multi-label classification of brain diseases from full-slice CT scans.
• To address the limitations of traditional slice-level analysis by incorporating dependencies between CT image slices.
• To reduce the manual labeling effort required for training deep learning models.

Main Methods:

• Proposed the Slice Dependencies Learning Model (SDLM), a novel deep learning architecture.
• The SDLM learns image features and slice dependencies from variable-length brain CT scan series.
• The model requires only the full-slice brain scan for disease labeling, simplifying the process.

Main Results:

• Evaluated on the CQ500 dataset (491 subjects, 1194 CT scan sets).
• Achieved a precision of 67.57%, recall of 61.04%, and F1 score of 0.6412.
• Demonstrated strong performance with an Area Under the ROC Curve (AUC) of 0.8934.

Conclusions:

• The SDLM offers a new approach for multi-label classification using full-slice brain CT scans.
• This architecture overcomes limitations of traditional slice-level classification methods.
• The model shows significant potential for improving multi-label disease detection in brain CT imaging.