A study of applying deep learning technology in medical images diagnosis for intracranial aneurysms
Intracranial aneurysms (IA) have relatively high prevalence among people. The rupture of IA is life-threating. Therefore, it is important to develop tools to realize automatic detection of IA and its rupture risk prediction. Currently, most tools for IA detection are based on 3-dimension (3D) geometry models, which need elaborated segmentation on medical images. This makes the currently available tools hard to be widely used by radiologists. In another aspect, most current tools for IA rupture risk prediction does not directly take medical images as input, or combine medical images with other important features, such as patient habits or hemodynamic characteristics of IA, as inputs. According to above review of literature, improving current tools for IA detection and rupture risk prediction is urgently needed. In this thesis, I developed a series of useful tools to assist radiologists in diagnosis of IAs. In Chapter 1, an introduction of this thesis is conducted. In Chapter 2, the main material and methods used in this thesis are presented. In Chapter 3, a tool was developed to distinguish IA vessels from normal vessels, which can be used to eliminate some false positives proposed by IA detection models. In Chapter 4, I developed a convolution neural network (CNN) model to realize the automatic detection of IA on computed tomography angiography (CTA) images. The sensitivity of the model achieved 91.8% with elapsed time less than 30s per case. The fast calculation, high sensitivity and simplicity of operation make it a promising tool to assist radiologists in IA detection. In Chapter 5, I developed a framework combined CTA images and patients’ habits as inputs to predict the rupture risk of aneurysms. The framework includes two parts: 3D CNN and logistic regression model. The 3D CNN takes original CTA images as input and calculates the CTA index for rupture risk. The logistic regression combines the CTA index and patients’ habits to predict the final rupture risk for IAs. By adding patients’ habits as inputs, the performance of IA rupture risk prediction can improve a lot. The sensitivity for ruptured IAs was 90% (95% CI: 55–100%) and the area under the receiver operating characteristic (AUROC) was 0.86 when using this framework, while the sensitivity was 75% (95% CI: 48–93%) when only taking CTA images as inputs. In another aspect, since hemodynamic characteristics also impact rupture risk, I also developed a CNN-FCN model that takes these characteristics as inputs to predict IA rupture risk in Chapter 6. The model takes original CTA images and hemodynamic characteristics as inputs. I calculated energy loss (EL), normalized max wall shear stress (NMWSS), mean wall shear stress (MWSS), low shear stress area ration and pressure loss coefficient (PLC) for all IAs in the dataset. With these hemodynamic characteristics, as well as original CTA images, the sensitivity, specificity and accuracy of the CNN-FCN model were 0.82 (95% CI: 0.70 – 0.91), 0.85 (95% CI: 0.71 – 0.94) and 0.84 (95% CI: 0.75 – 0.90), respectively. The AUROC and the area under the precision-recall curve (AUPRC) were 0.87 and 0.91, respectively. In Chapter 7, the main conclusions and outlooks for this thesis are summarized. In summary, this thesis developed a series of models to assist radiologists in IA detection and rupture risk prediction. These tools can help radiologists improve working efficiency in clinical environment. A few important features to predict rupture risk of IA were confirmed in this study.