An Efficient Framework to Detect Intracranial Hemorrhage Using Hybrid Deep Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Intracranial hemorrhage (ICH) is a critical condition requiring rapid diagnosis.
• Accurate classification of multiple ICH types (epidural, intraparenchymal, intraventricular, subarachnoid, subdural) is challenging.
• Current diagnostic methods rely on expert interpretation of CT scans.

Purpose of the Study:

• To develop and evaluate a multi-label classification model for detecting and classifying six types of ICH.
• To improve the accuracy and efficiency of ICH diagnosis using deep learning.
• To address the challenge of simultaneous multiple hemorrhages in patients.

Main Methods:

• A hybrid deep learning model combining Convolutional Neural Network (CNN) and Long-Short Term Memory (LSTM) (Conv-LSTM) was developed.
• A Systematic Windowing technique was integrated with the Conv-LSTM model.
• Experiments were conducted using the publicly available RSNA dataset for ICH detection.

Main Results:

• The proposed Conv-LSTM model demonstrated high performance metrics: 93.87% sensitivity, 96.45% specificity, 95.21% precision, and 95.14% accuracy.
• The model effectively identifies the presence of hemorrhage and classifies its type(s).
• The F1 score results surpassed those of existing deep neural network-based algorithms.

Conclusions:

• The hybrid Conv-LSTM model offers a promising solution for accurate multi-label ICH classification from CT scans.
• This approach can aid clinicians in faster and more precise diagnosis of various intracranial hemorrhages.
• The model's performance suggests potential for integration into clinical workflows for improved patient outcomes.