{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/ternarynet-faster-deep-model-inference","title":"TernaryNet: Faster Deep Model Inference without GPUs for Medical 3D Segmentation using Sparse and Binary Convolutions","arxiv_id":"1801.09449","date":"2018-01-29","proceeding":null,"authors":["Mattias P. Heinrich","Max Blendowski","Ozan Oktay"],"abstract":"Deep convolutional neural networks (DCNN) are currently ubiquitous in medical\nimaging. While their versatility and high quality results for common image\nanalysis tasks including segmentation, localisation and prediction is\nastonishing, the large representational power comes at the cost of highly\ndemanding computational effort. This limits their practical applications for\nimage guided interventions and diagnostic (point-of-care) support using mobile\ndevices without graphics processing units (GPU). We propose a new scheme that\napproximates both trainable weights and neural activations in deep networks by\nternary values and tackles the open question of backpropagation when dealing\nwith non-differentiable functions. Our solution enables the removal of the\nexpensive floating-point matrix multiplications throughout any convolutional\nneural network and replaces them by energy and time preserving binary operators\nand population counts. Our approach, which is demonstrated using a\nfully-convolutional network (FCN) for CT pancreas segmentation leads to more\nthan 10-fold reduced memory requirements and we provide a concept for\nsub-second inference without GPUs. Our ternary approximation obtains high\naccuracies (without any post-processing) with a Dice overlap of 71.0% that are\nstatistically equivalent to using networks with high-precision weights and\nactivations. We further demonstrate the significant improvements reached in\ncomparison to binary quantisation and without our proposed ternary hyperbolic\ntangent continuation. We present a key enabling technique for highly efficient\nDCNN inference without GPUs that will help to bring the advances of deep\nlearning to practical clinical applications. It has also great promise for\nimproving accuracies in large-scale medical data retrieval.","url_abs":"http://arxiv.org/abs/1801.09449v1","url_pdf":"http://arxiv.org/pdf/1801.09449v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"ternarynet-faster-deep-model-inference","repo_url":"https://github.com/mattiaspaul/TernaryNet","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"3d-medical-imaging-segmentation","task_name":"3D Medical Imaging Segmentation"},{"task_slug":"diagnostic","task_name":"Diagnostic"},{"task_slug":null,"task_name":"GPU"},{"task_slug":"open-question","task_name":"Open-Ended Question Answering"},{"task_slug":"pancreas-segmentation","task_name":"Pancreas Segmentation"},{"task_slug":"retrieval","task_name":"Retrieval"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":null,"mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}