| Aneurysms of left atrium are rare abnormalities.They can be congenital or acquired Whereas a true congenital aneurysm presents as isolated pathology,inflammatory or de-generative processes involving the endocaordium are associated with the acquired type High-resolution cardiac CT represents a unique tool to asses subtle anatomic cardiac vari-ants,which is useful for ruling out the differential diagnoses of aneurysm of the acquired typeIn this thesis primarily try to research three problem which all are based on the subject of the deep learning method in left atrial aneurysm detection.Firstly we want to figure out what are the nature,properties,and usage of the multi-scale of anchors technology in the state-of-art object detection framework(like RetinaNet).By reviewing the existing works and by conducting experiments of our own,we make a systematic presentation of what the multi-scale anchors are and how they works.We leverage the efficient CNN architecture of RetinaNet for CADe system.However,an unusual gap between the validation set and the cross-validation set takes our curiosity As we explore further,raises the issue of how we human assess annotations with proper contextual information for CNN.We try to tackle this problem by proposing a simple soft thresholding model,which is found on the assumption that there is a negative correlation between aneurysm size and the annotation scale that is agreeable to CNN.The ultimate results,with mAP(Mean Average Precision)increasing from 0.699 to 0.735 in K-fold cross validation experiments,have justified our explanation and solution.In the end,we move to the area to 3D imaging processing.In order to create a 3D dataset from our existing 2D dataset of aneurysms directly.At first,We do mathematically mod-eling of an aneurysm consists of 2D records.In addition we also develop an self-adaptive k-means clustering algorithm which is based on the lathematical model.The visual-ization of our generated 3D datset show our clustering results matching our expectation.Afterwards,we adapt a two-stage 3D convolutional framework newly proposed on MIC-CAL to our aneurysm detection problem.The results of the adaptive model on our 3D dataset are:recaL1=0.8846,FPs(False Positives per Scan)=28.52. |
