This poster will be presented by Luciano Lucas, PhD on Monday December 10th from 12:00 - 1:30 PM. Poster number P1898.
The technology presented in this poster is a part of our latest release, Aivia Cloud. Request a demo to try it out: https://www.aivia-software.com/demo
Machine learning (ML) for imaging applications has grown dramatically in recent years. Deep learning (DL) powered approaches have been shown to improve image resolution, increase signal to noise ratio (SNR),automatically identify objects of interest, and predict the localization of various fluorescently labelled subcellular features from microscopy images. While the number of scientific publications describing new ways to use ML / DL for imaging applications continues to rise sharply, very few biologists and microscopists have adopted these approaches. Four practical hurdles prevent practical adoption for all but ML/DL experts:
A. No single software covers the entire workflow, forcing users to master multiple specialized apps, which requires transferring and reformatting large volumes of data between apps.
B. DL model training requires expert knowledge to finetune.
C. DL model training requires multiple high performance computers.
D. Ground truth (GT) data for DL training is often difficult to obtain.
To address these issues, we have developed a platform which supports end-to-end DL workflow: 1) Creation of GT 2) Training DL models 3) Apply trained DL models 4) Perform image segmentation 5) Review original data and results 6) Edit results and 7) Export results.
We have created several innovative enhancements at different steps to augment the end to end workflow. In step 1, a new ML enabled tool provides 8x speed up for creating GT annotations in comparison to conventional methods. For steps 2 through 5, we have developed Aivia Cloud, a novel custom client leveraging virtual machine instances using Google Cloud Platform. Steps 2 and 3 make use of connecting Aivia (desktop ) to Aivia Cloud and allow for the easy setup of DL jobs. We found that training and applying using an Aivia Cloud instance with a single nVidia V100 GPU was up to 10x faster as compared to Aivia on a local machine with an nVidia GTX 1080ti GPU. Step 4 can be accelerated up to 20x vs Aivia. Remote access via Aivia Cloud in step 5 avoids the need to download the results for local inspection. Step 6 makes use of a new real-time 3D mesh editing tool which reduces the editing time by up to 5x vs standard methods.
Our novel hybrid approach removes hurdles A through C and greatly mitigates D. These are the major bottlenecks that block biologists and microscopists from taking advantage of DL powered solutions for their applications.
Future work includes the cataloging and benchmarking of popular DL models for arrange of applications as well as integrating the DL training / applying app 2) and 3) into Aivia’s ecosystem. We propose Aivia / Aivia Cloud as a streamlined end-to-end pipeline for image visualization and analysis (including deep learning) which can easily be used by any scientist.
Authors: H.Sasaki, J.Stansberry, M.Jones, S.McElroy, C.McBride, C.Birnbaum, B.Graff, T.Cheng, Z. Kenyon, T.Lucas, M.Hsieh, T.Phan, H.Lai, C.Huang, J.S.Lee, L.A.Lucas
Authors are a part of DRVISION Technologies LLC, Bellevue, WA