- Information
- AI Chat
Practical - Diagnosis of heart disease via cnns
Convolutional Neural Networks for Visual Recognition (CS 231n)
Stanford University
Recommended for you
Students also viewed
- CS231n 2016-2017
- Practical - Imagenet classification with deep convolutional neural networks
- Summary - Gradient based learning applied to document recognition
- Practical - Multi-view 3d pose estimation from single depth images
- Summary - Real-time depth estimation from 2d images
- Practical - Random search for hyper-parameter optimization
Related documents
Preview text
Diagnosis of Heart Disease via CNNs Kaicheng Wang Stanford University Yanyang Kong Stanford University Abstract volume VD from MRI time series data in different axis views (the planes of slice). Our project predicts volume of heart 2D MRI measurement. Combining VGG and selftrained networks, we build our Convolutional Neural Networks (CNNs) for prediction. Specific preprocessing methods are designed for our messy data. CNNs in various depths and regularization strength are tried for best validation result. In this Kaggle Challenge contest, our model beats the baseline CNN structure written Marko Jocic (a) volume Figure 1: Plots of of volumes in one cardiac circle, circled in red curve 1. Introduction Using MRI data to measure cardiac (VS ) and (VD ) volumes (i., the size of one chamber of the heart at the beginning and middle of each heartbeat) and then deriving the ejection fraction (EF) of heart is a standard process to assess the squeezing ability. Declining EF is a key indicator of heart disease. However, the current process is manual and slow. The cardiologist could spend up to 20 minutes with one patient. Considering the huge amount of patients with potential heart failure, quick automatic measurement will help doctors to diagnose heart conditions more efficiently. 1 EF 100 (b) volume 1. Problem Statement Given one several (10 axes in average) MRI measurements in different axes (sax 5,, 2ch 12, 4ch 32, etc), we are going to predict his or her left enddistole volume and volume in one cardiac cycle. Each measurement contains 30 times series images in one cardiac cycle. We aim to build two deep CNN regression models to predict these two volumes separately. The overview of our problem is shown in fig VD VS VD Convolutional Neural Networks (CNNs) have been proved remarkably effective on neuroimaging data. As a powerful visual model, CNNs can yield many interesting hierarchies of features, which can be used to classification and segmentation. Contemporary CNN models like have been used and transferred to learn representations weights in many different tasks. In our project, we are going to apply deep Convolutional Neural Networks to predict the volume VS and Figure 2: Using 30 MRIs during one cardiac cycle from different axis views to predict VS and VD 1 This project is the Second Annual Data Science Bowl Challenge lem: 1 1. Challenges 3. Data and Preprocess In this problem, variation in heart shape and image quantity of patients makes automated quantification of volume challenging. In the training dataset, we have a diverse representation of cases. Patients form different hospitals may be measured MRI in various axes. Besides, normal and abnormal images of hearts are mixed together. In an extreme case, we have different poses and directions even in one cardiac cycle. Thus these images are highly distinct from each other. We need a robust model to validate and automate the manual measurement of ejection fraction. We will examine hundreds of cardiac MRI images in DICOM format for each patient. This dataset was compiled the National Institutes of Health and National Medical Center. It is an order of magnitude larger than any cardiac MRI data set released previously. We only utilize pixel information of DICOM file. 3. Details of Data In the training dataset, we have 500 patients undergoing about 10 experiments (measurements) from different axes planes. Each experiment observes one slice of heart, which leads to 30 images across the cardiac cycle. Different experiments are acquired from separate breath holds. This is important since the registration from slice to slice is expected to be imperfect. Besides, there are another 200 patients in the validation dataset. In short, We have about raw training images and 200 10 30 raw validation images. 2. Related Work Deep learning, especially Convolutional Neural Networks have been applied to medical imaging recognition in recent years. Sergey Plis, et al. applied deep neural networks to learn structural and functional brain imaging data and showed that deep learning methods are able to learn physiologically important representations and detect latent relations in neuroimaging data. Payan et al. used 3D convolutional neural network to predict disease based on brain MRI images. They first initialized the filters for CNN sparse and then built a CNN whose first layer using the filters learned with Their model proved to be a success in discriminating between healthy and diseased brains. However, their neural network model was too simple and they go on to try more sophisticated neural network models. Liu and Shen also applied deep CNN models on raw MRI images. They found the regions of interest (ROI) that may be correlated with disease, which spares the effort of manual ROI annotation process. Brebisson et al. constructed a deep convolutional neural network called SegNet for anatomical brain segmentation. In cardiac study area, N Kannathal, et al. have implemented neural networks for classification of cardiac patient states using electrocardiogram (ECG) signal. Yaniv Bar et al. used a CNN trained ImageNet and obtained area under curve (AUC) of 0 for Right Pleural Effusion detection, 0 for Enlarged heart detection, 0 for classification between healthy and abnormal chest which shows that deep learning with large scale image database may be sufficient for general medical image recognition tasks. In terms of our specific problem, there is a baseline CNN structure written Marko Jocic He simply treated all images equally and trained a ConvNet with 6 and 2 Each group of 30 images is an input and the corresponding volume is output. This model is easy to compute but not precise. We will show how we beat the baseline model treating images differently and more accurate models. 3. Preprocessing Notice that the amount of measurements of cardiac MRI varies from patient to patient (not necessarily 10). These images are taken from different views, or axis planes in medical imaging terminology, like 2ch 16, 4ch 17, sax 5, sax 6, sax 7, sax 8,etc. Each patient has different planes for his or her heart. According to ch represents left ventricular long axis acquisition planes. The 2 ch (fig and 4 ch views are used to visualize different regions of the left atrium, mitral valve apparatus. sax represents short axis acquisition planes(fig These stacks are oriented parallel to the mitral valve ring, and are acquired regularly spaced from the cardiac base to the apex of the heart. (a) short axis stack (b) view (c) view Figure 3: Views of sax, 2 ch, 4 ch. Since sax views are excellent in volumetric measurements, as shows, we divide these planes into four regions, with three regions for continuous sax stacks, called correspondingly and another region for ch views called ch. The corresponds to the first third of sax views of that patient, corresponds to the second third of the sax views and corresponds to the last third of the sax views. Region ch represents for and views, which are additional views and less important than sax views. 2 Model VGG and network are different from each other. We optimize our model according to the evaluation result. Details of models are introduced in Section 4. The reason for building such models are explained in Section 5. 1 2 FC(128, 256, 600, 1)(to be trained) 3 16 . FC(128, 256, 600, 1)(to be trained) 4 15 1 3, 3)(to be trained) . FC(128, 256, 600, 1) (to be trained) 5 15 1 3, 3)(to be trained) FC(128, 256, 600, 1) (to . be trained 6 15 1 3, 3)(to be trained) dropout(0) FC(128, 256, 600, 1) (to . be trained 7 15 1 3, 3)(to be trained) FC(128, 256, 600, 1) (to . be trained 8 15 1 3, 3)(to be trained) dropout(0) FC(128, 256, 600, 1) (to be trained 4. Evaluation Since our outputs are 2 volumes, VS and VD , we use Root Mean Squared Error (RMSE) as our loss function. We implement CNNs in Keras with Adam optimizer. Based on our volume prediction and loss 2 , we first calculate CDF (cumulative probability distribution) for VS and VD fitting normal distribution with mean as predicted class value and variance as loss N 2 ). Intuitively the predicted distribution curve will be sharper when we have lower loss, since we are more confident about our prediction. Then our model is evaluated on the Continuous Ranked Probability Score (CRPS) as mentioned in Kaggle, see N 599 1 XX (P (y n) H(n Vm ))2 600N where H(x) is the Heaviside step function (H(x) 1 for x 0 and 0 otherwise). Structure . . FC(128, 256, 600, 1)(to be trained) Table 1: The structures of 8 models in our experiments removes the last in VGG and adds one to the network. to keep the same structure of VGG as in and the only difference between them is the strength of regularization. The schematic diagrams of all models are showed in the table Figure 6: plot of predicted distribution, with error measured as the area of green region. 4. Models Our model is divided into 2 parts. As for the pretrained VGG part, we compare results of 3 models (Model1, the depths of which increase as more are added. As for the network part after VGG, we explored optimal parameters of 1 convlayer and 2 (FC) layers to Model8). See the table for more details. 8 models are experimented for VS and VD respectively. Since performs better than and Model2 in either cases, to are built based on In the end, we choose the best model for predicting VS and VD independently according to CRPS on validation data. to utilize same FClayers while adding more pretrained Based on 5. Results and Discussion We run our models on Stanford Rye01 with configuration as 8 core (2x E5620) cpu, 48GB ram, 250GB local disk, 6x C2070, Ubuntu 13 and CUDA 6. In each model, we use mini batch gradient descent to train the weights and the batch size is 100. We train the models over 400 iterations and get the CRPS for 8 models in the table We choose the optimal systolic model and diastolic model CRPS on validation dataset. According to the does perform the best in first 3 models in either cases, which implies that deeper lay4 Figure 7: CRPS for 8 models ers are better for our transferable learning. This indicates some similarities between our dataset and ILSVRC dataset, where the VGG is trained. At present, only 2 are trained in our models. We then take a further step to add a trainable on top of our network at the same time we remove the last in pretrained VGG just to be fair. That is how is built. We notice overfitting problem in the first 4 models. adding dropout and choosing different regularization strength, we experiment 4 more models to fight against overfitting. Since performs best in the first 4 models, we set to inherit identical structure from According to the final results, We choose to predict systolic volume VS and to predict diastolic volume VD . Notice that for every model, we trained endsystolic and model separately. In other words, they share the same structure but are different in weights. In experiment, the appropriate learning rates differ from model to model. We tune all 16 models manually. Endsystolic models are trained at learning rate 1e 5 imately. learning rates are usually two magnitudes higher. We present these two best learning process in the first 100 iterations in Figure 8: RMSE of systole and diastole models in 100 iterations. 5 S. M. Plis, D. R. Hjelm, R. Salakhutdinov, and V. D. Calhoun. Deep learning for neuroimaging: a validation study. arXiv preprint arXiv:1312, 2013. K. Simonyan and A. Zisserman. Very deep convolutional networks for image recognition. arXiv preprint arXiv:1409, 2014. C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proceedings of the Conference on Computer Vision and Pattern Recognition, pages 2015. 7
Practical - Diagnosis of heart disease via cnns
Course: Convolutional Neural Networks for Visual Recognition (CS 231n)
University: Stanford University
- Discover more from:
Recommended for you
Students also viewed
- CS231n 2016-2017
- Practical - Imagenet classification with deep convolutional neural networks
- Summary - Gradient based learning applied to document recognition
- Practical - Multi-view 3d pose estimation from single depth images
- Summary - Real-time depth estimation from 2d images
- Practical - Random search for hyper-parameter optimization