Intra-operative Optimal Flow Diverter Selection for Intracranial aneurysm treatment

TL;DR

This study explores a machine learning approach using Angiographic Parametric Imaging to optimize flow diverter selection during intracranial aneurysm treatment, aiming to improve intra-operative decision-making.

Contribution

It introduces a novel ML-based method utilizing API parameters for intra-operative FD sizing guidance, addressing limitations of CFD-based methods.

Findings

ML approach achieved AUROC of 0.72 for predicting undersized FDs

Single API parameters were ineffective with AUROC ~0.5

Method shows potential for intra-operative aneurysm treatment guidance

Abstract

During intracranial aneurysm (IA) treatment with Diverters (FDs), the device/parent artery diameters ratio may influence the ability of the device to induce aneurysm healing response. Oversized FDs are safer to deploy but may not induce enough hemodynamic resistance to ensure aneurysm occlusion. Methods based on Computational Fluid Dynamics (CFD) could allow optimal device selection but are time-consuming and inadequate for intra-operative guidance. To address this limitation, we propose to investigate a method for optimal FD selection using Angiographic Parametric Imaging (API) and machine learning (ML). We selected 128 pre-treatment angiographic sequences of IAs which demonstrated full occlusion at six months follow-up. For each IA, we extracted five API parameters from the aneurysm dome and normalized them to the feeding artery corresponding parameters. We dichotomized the dataset based on the FD/ proximal artery diameter ratio as undersized, if the ratio<1 or if multiple FDs were used and oversized otherwise. Single API parameter and ML analysis were used to determine whether API parameters could be used to determine the need for FD under-sizing (i.e., increased flow resistance). Classification accuracy was assessed using area under the receiver operator characteristic (AUROC). In total we identified 51 and 77 cases for the undersized and oversized cohorts respectively. Single API parameter analysis yielded an inadequate AUROC ~0.5 while machine learning using all five API parameters yielded and AUROC of 0.72.