Deep learning and machine learning predictive models for neurological function after interventional embolization of...

Area of Science:

• Neurosurgery
• Medical Imaging
• Artificial Intelligence

Background:

• Intracranial aneurysms pose significant risks for postoperative neurological complications.
• Accurate prediction of postoperative Hunt-Hess grade is crucial for patient management and prognosis.

Purpose of the Study:

• To develop a predictive model for postoperative Hunt-Hess grade in intracranial aneurysm patients.
• To integrate radiomics and deep learning using preoperative CTA data.
• To enhance clinical decision-making and improve assessment of neurological function.

Main Methods:

• Retrospective analysis of 101 patients undergoing aneurysm embolization.
• Extraction of 851 radiomic and 512 deep learning features from CTA images.
• Application of feature selection techniques and construction of radiomics (RSM), deep learning (DLM), and fusion (DLRSCM) models using ensemble learning (Stacking).

Main Results:

• The deep learning-radiomics feature fusion model (DLRSCM) achieved the highest performance with an AUC of 0.968 and MCC of 0.820.
• The deep learning model (DLM) showed an AUC of 0.959 and MCC of 0.815.
• The radiomics model (RSM) achieved an AUC of 0.935 and MCC of 0.793, with stacked ensemble models outperforming base algorithms.

Conclusions:

• The integration of radiomics and deep learning effectively predicts postoperative Hunt-Hess grade in intracranial aneurysm patients.
• This combined approach offers significant value for early identification of neurological complications.
• The findings support enhanced clinical decision-making for improved patient outcomes.