TernaryNet: faster deep model inference without GPUs for medical 3D segmentation using sparse and binary convolutions

Published in

International Journal of Computer Assisted Radi...(2018)

External Links:

Generate Graph DownloadPDF

TL;DR

This work proposes a new scheme that approximates both trainable weights and neural activations in deep networks by ternary values and tackles the open question of backpropagation when dealing with non-differentiable functions and presents a key enabling technique for highly efficient DCNN inference without GPUs.

Abstract

PurposeDeep convolutional neural networks (DCNN) are currently ubiquitous in medical imaging. While their versatility and high-quality results for common image analysis tasks including segmentation, localisation and prediction is astonishing, the large representational power comes at the cost of highly demanding computational effort. This limits their practical applications for image-guided interventions and diagnostic (point-of-care) support using mobile devices without graphics processing units (GPU).MethodsWe propose a new scheme that approximates both trainable weights and neural activations in deep networks by ternary values and tackles the open question of backpropagation when dealing with non-differentiable functions. Our solution enables the removal of the expensive floating-point matrix multiplications throughout any convolutional neural network and replaces them by energy- and time-preserving binary operators and population counts.ResultsWe evaluate our approach for the segmentation of the pancreas in CT. Here, our ternary approximation within a fully convolutional network leads to more than 90% memory reductions and high accuracy (without any post-processing) with a Dice overlap of 71.0% that comes close to the one obtained when using networks with high-precision weights and activations. We further provide a concept for sub-second inference without GPUs and demonstrate significant improvements in comparison with binary quantisation and without our proposed ternary hyperbolic tangent continuation.ConclusionsWe present a key enabling technique for highly efficient DCNN inference without GPUs that will help to bring the advances of deep learning to practical clinical applications. It has also great promise for improving accuracies in large-scale medical data retrieval.

TernaryNet: faster deep model inference without GPUs for medical 3D segmentation using sparse and binary convolutions

Published in

International Journal of Computer Assisted Radi...(2018)

External Links:

Generate Graph DownloadPDF

TL;DR

Abstract

References42 items

Towards Image-Guided Pancreas and Biliary Endoscopy: Automatic Multi-organ Segmentation on Abdominal CT with Dense Dilated Networks

Hashing with Residual Networks for Image Retrieval

BRIEFnet: Deep Pancreas Segmentation Using Binary Sparse Convolutions

Robust Abdominal Organ Segmentation Using Regional Convolutional Neural Networks

Corticomuscular Coherence With Time Lag With Application to Delay Estimation

HashNet: Deep Learning to Hash by Continuation

Variational Dropout Sparsifies Deep Neural Networks

A Fixed-Point Model for Pancreas Segmentation in Abdominal CT Scans

Faster Population Counts Using AVX2 Instructions

Trained Ternary Quantization

Multi-organ Segmentation Using Vantage Point Forests and Binary Context Features

Dilated Convolutional Neural Networks for Cardiovascular MR Segmentation in Congenital Heart Disease

Densely Connected Convolutional Networks

Local Binary Convolutional Neural Networks

Spatial Aggregation of Holistically-Nested Networks for Automated Pancreas Segmentation

V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation

Evaluation of Six Registration Methods for the Human Abdomen on Clinically Acquired CT

XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks

Evaluation of body-wise and organ-wise registrations for abdominal organs

Binarized Neural Networks

Multi-Scale Context Aggregation by Dilated Convolutions

Efficient multi-atlas abdominal segmentation on clinically acquired CT with SIMPLE context learning

DeepOrgan: Multi-level Deep Convolutional Networks for Automated Pancreas Segmentation

Sparse Convolutional Neural Networks

U-Net: Convolutional Networks for Biomedical Image Segmentation

Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Towards Large-Scale Histopathological Image Analysis: Hashing-Based Image Retrieval

Fully convolutional networks for semantic segmentation

Edge- and Detail-Preserving Sparse Image Representations for Deformable Registration of Chest MRI and CT Volumes

BRIEF: Computing a Local Binary Descriptor Very Fast

Toward Precise Pulmonary Nodule Descriptors for Nodule Type Classification

Recognizing linked events: Searching the space of feasible explanations

Efficient sparse coding algorithms

Emergence of simple-cell receptive field properties by learning a sparse code for natural images

Integrating geometric configuration and appearance information into a unified framework for anatomical landmark localization

Medical Image Computing and Computer-Assisted Intervention − MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part II

Ternary weight networks. In: NIPS workshop on efficient methods for deep neural networks (EMDNN)

Ran El-Yaniv, and Yoshua Bengio

Multi-modal Multi-Atlas Segmentation using Discrete Optimisation and Self-Similarities

Learning Semantic Image Representations at a Large Scale

Convolutional Deep Belief Networks on CIFAR-10

: Faster Deep Model Inference w