Aneurysm detection in 3D cerebral angiograms based on intra-vascular distance mapping and convolutional neural networks

Intracranial Aneurysm Detection from 3D Vascular Mesh Models with Ensemble Deep Learning

“…Our results are superior in some of the metrics, depending on the compared study, although a more global and direct comparison is hindered by the differences in data sets and evaluation criteria. Our algorithm obtain similar or superior results to both hand-crafted feature-based traditional methods and other deep leaning methods [12,16,5,10,4], which requires more homogeneous data formats. Our results indicate that (1) conformal parameterization method provides a valid mesh representation, (2) pre-trained and fine-tuned V3 model can effectively transfer the image representation ability learned on the ImageNet dataset to characterize the parameterization method model, (3) an adaptive ensemble of these images has superior ability in identifying aneurysms in cerebrovascular mesh models.…”

“…1. The algorithm has four steps: (1) constructing the conformal mapping from input cerebrovascular mesh model S to the sphere-like surface S ′ and then conformal mapping from S ′ to planar flat-torus τ with seamless counter-clockwise covers of S ′ (2) building a MMEN architecture for 3D mesh model based classification (3) training the MMEN with the input dataset {S i } described with d-dimensional features until convergence, and (4) detecting IAs in each model based on the labels of aneurysms. Problem formulation Each 3D cerebrovascular structure can be modeled as a series of connected tubular-like 3D branches model.…”

“…More recently, deep learning, particularly the deep convolutional neural network (DCNN), has become the most successful technique in IAs detection and it provides a unified framework for joint feature extraction and detection. Jerman [4] first utilized 7 layer 2D CNN to detect IAs with intra-vascular distance map. Nakao [10] utilized 6 layer 2D CNN on 9 direction MIP images of each cube VOI.…”

Aneurysm Identification in Cerebral Models with Multiview Convolutional Neural Network

“…Since the use of CNNs has been successful in computer vision and image processing, many studies have examined aneurysm detection in medical images, such as MRA or 3DRA using a CNN. Jerman [5] used a Hessian-based filter to enhance spherical and elliptical structures such as aneurysms and attenuate other structures on the angiograms. Next, they boosted the classification performance using a 2D CNN trained on intravascular distance maps computed by casting rays from the preclassified voxels and detecting the first-hit edges of the vascular structures.…”

“…The inputs of the network were 2D images generated from volumes of interest of the MRA images by applying a mixed-integer programming algorithm. The network architecture they used was not very deep: 4 convolution layers in one [5] and 2 convolution layers in the other [10]. More adjustable parameters (weights and bias) correspond to greater freedom of adjustment and a better approximation effect.…”