Research On Cerebral Aneurysms Detection In CTA Images Based On 3D Convolutional Neural Network

Cerebral aneurysms are tumor-like objects formed by local bulging or swelling of cerebral arteries.The corresponding location of the artery wall is thin,and the risk of rupture is high.Once ruptured,it is easy to cause intracranial hemorrhage and threaten life.At present,the commonly used medical technique for rapidly diagnosing aneurysms is computed tomography angiography(CTA).While this technology brings the convenience of rapid imaging(widely used in emergency departments),it also contains complex background noise,making it more difficult for radiologists to read.Radiologists often need to spend a long time diagnosing aneurysms in CTA images.It is easier for some less experienced radiologists to miss an aneurysm,which will delay the precious time for treatment.Therefore,the research on the automatic detection methods of cerebral aneurysms with CTA images is of great significance to assist radiologists in quickly diagnosing the condition and saving lives.The detection of cerebral aneurysms in CTA images faces the problems of lack of publicly diverse data sets and the tendency to miss detection and false alarm.The aneurysm data in the actual scene often contains multi-center,multi-device,multi-location,multi-type,and multi-scale representations.Currently,there are few publicly available CTA image datasets with these diverse representations that are close to real scenarios.The problem of easy to miss and false detection is reflected in the small size of cerebral aneurysms,which is not easy to recall as well as the complex background of CTA images,which is easy to introduce false-positive prediction.Considering the problems of cerebral aneurysm detection tasks in CTA images,this thesis carried out the following research:First of all,in response to the lack of publicly available diverse data sets,this thesis collected CTA images with cerebral aneurysms and radiologists’ professional annotations from two hospitals.Based on this,this thesis proposes a baseline model CANet for cerebral aneurysm detection.Compared with the current multi-step aneurysm detection algorithm,CANet directly inputs 3D CTA image data for end-to-end cerebral aneurysm detection without complicated pre-processing and post-processing procedures which has higher practical value.The experimental results show that CANet has achieved better detection performance than other advanced algorithms on two actual data sets.Secondly,to solve the problem of easily missed detection and false detection,this thesis proposes a novel change feature aggregation module.In this module,this thesis imitates the radiologists’ technique of reading the CTA images,i.e.,diagnosing cerebral aneurysm with the help of change features between the current slice and its adjacent slice.This module fully considers the change feature between CTA image slices.It can be flexibly integrated into the existing cerebral aneurysm detection algorithm,which can effectively improve its detection performance with less increase in the number of parameters.By integrating the change feature aggregation module into CANet,this thesis proposes an improved version of the cerebral aneurysm detection algorithm CANet++.The results of experiments show that CANet++ has achieved state-of-the-art on two cerebral aneurysm data sets thanks to the aggregation of inter-slice change features.In summary,this thesis collects actual data sets that do not require much manual preprocessing and have diverse aneurysm attributes.Based on the data sets,this thesis proposes an end-to-end aneurysm detection algorithm with a relatively simple process.Besides,this thesis proposes an inter-slice change feature aggregation module to improve the performance of cerebral aneurysm detection further.