On-device Training for Breast Ultrasound Image Classification

2020 10th Annual Computing and Communication Workshop and Conference (CCWC)
Date of Conference: 06-08 January 2020
Conference Location: Las Vegas, NV, USA

Transcript

1. 1 On-device Training for Breast Ultrasound Image Classification Dennis Hou, Rutgers University Raymond Hou, Northshore University Hospital Janpu Hou, Applied Data Research January 7, 2020
2. 2 On-device Training for Breast Ultrasound Image Classification •Dennis Hou, Raymond Hou, Janpu Hou •Published 1 January 2020 •Computer Science •2020 10th Annual Computing and Communication Workshop and Conference (CCWC) Most on-device AI pre-trained a neural network model in cloud-based server then deployed to edge device for inference. On-device training not only can build personalized model, but also do distributed training like federated learning to train accurate models from scratch using small updates from many devices. In this work, we implement the semi-supervised convolutional neural network based on successive subspace learning and use a dataset of breast ultrasound (BUS) images to demonstrate a proof of concept of true on-device training. An important advantage of such network is that we can extract the key feature vectors with CNN network architectures without the need of backpropagation computation made it suitable for portable ultrasound. It can acquire the ultrasound image and train the CNN classifier on the portable device without cloud-based server. We evaluate the model by using a set of BUS images that includes benign and malignant breast tumors. We obtain 94.8% accuracy with this study and demonstrate the applicability of the proposed on-device training model to improve the diagnosis of BUS images.
3. 3 Outlines: Introduction • Breast Cancer Screening • Community Health Center Methodology • Successive Subspace Learning • Ensemble Classifier Results and Discussion
4. 4 Breast cancer screening Affordable Care Act makes breast cancer screening free • But not for uninsured or under-insured Second leading cause of women's death from cancer. • have a 98% chance of surviving if detected early
5. 5 Community Health Center mobile unit goes to: • underserved areas of our state(low budget) • geographically isolated areas(no internet) • those who are uninsured(need free service) Mobile mammography unit
6. 6 Problem Statement • There is a need for low budget breast ultrasound screening tool for un-insured and under-insured women. • Portable unit without internet connection. • Personalized on-device training for follow up medical check
8. 8 Mobile Ultrasound Embedded System Breast Ultrasound Classifier Personalized On-Device Training
9. 9 Outlines: Introduction • Breast Cancer Screening • Community Health Center Methodology • Successive Subspace Learning • Ensemble Classifier Results and Discussion
10. 10 • Convolution • Max pool • Non-linear Activation Filter Weights to be Learned from D ata LeCun A Standard CNN Architecture
11. 11 Spectral Spatial Calculate the variance of each patch and remove the small-variance patches. Image Submatrices Covariance Matrix of Image Submatrices Input Image • Convolution • Max pool A Saak Transform Architecture
12. 12 Obtain transform kernels with Karhunen–Loève Transform Covariance Matrix of Image Submatrices F - F Kernel Augmentation = lSaak F - F Saak Coefficients => Feature Map Image Submatrices • Non-linear Activation = lKLT F F A Saak Transform Architecture
13. 13 64x64x3 32x32x8 16x16x12 8x8x16 4x4x24 2x2x32 • Saak Transform: Convolution • Remove low variance: Max pool • ReLU + Augmented Kernels : Non- linear Activation Data-driven Filter Weights A Saak Transform Architecture
14. 14 1024 128 32 Features Classes A Multilayer Perceptron (MLP) Architecture
15. 15 1024 Clusters With Cluster labels Features With Class labels K-Means Clustering (Original Class Labels + New Cluster Labels) Pseudo-labels K-Mean Clustering
16. 16 128 Clusters With Cluster labels K-Means Clustering (Original Class Labels + New Cluster Labels) Pseudo-labels Pseudo-labels LSR1 Least-Squared Regression (LSR)
17. 17 Features Classes LSR1 LSR 2 LSR 3 Multi-stage label-guided LSRs
18. 18 Outlines: Introduction • Breast Cancer Screening • Community Health Center Methodology • Successive Subspace Learning • Ensemble Classifier Results and Discussion
19. 19 Benign Malignant Breast Ultrasound Images
20. 20 Output Classification Input Ultrasound Image Battery Power Breast Ultrasound Classifier
21. 21 Vary filter volume in Convolution Layers Testing accuracy increased
22. 22 Vary node number in 1st Fully Connected Layer Testing accuracy increased
23. 23 Energy % vs Testing Accuracy
24. 24 Summary: Simple Methodology applied to Breast Ultrasound Image Screening • No stand framework TensorFlow or PyTorch required • No internet connection required • Easy to build Ensemble Classifier
25. 25 Who Has Cited This Publication?