Deep learning based models like ResNet, DenseNet, InceptionNet, etc. have revolutionized radiologic image processing. Large memory and computational requirements make deployment of these models on low-powered Acorn RISC Machine (ARM) devices challenging. We worked on developing low-powered and compressed versions of deep models focused on chest X-rays.

Explored compression techniques include i) Quantization: model parameters conversion to low-precision data-types like i-a) int8 or i-b) float16 by rounding off scaled parameters. Run-time selection of scaling factor implies i-c) dynamic quantization. ii) Pruning removes redundant model links with minimal performance reduction. iii) Hybrid technique combines pruning and quantization. 

We used i) RSNA pneumonia detection dataset (30000 X-rays, binay classification) with i-a) DenseNet201 and i-b) InceptionNet-V3, and ii) Chest X-ray14 dataset (112120 X-rays, 14 findings, multi-class classification) with ii-a) ChexNet.

Dynamic quantization performs the best overall. It reduces InceptionNet-V3 size from 251MB to 21MB with 0.1% performance improvement. CheXNet size reduces from 29.4MB to 7.3MB while keeping performance stable. DenseNet gains the most size reduction under int8 quantization (211MB to 18MB) with 7% performance reduction while float16 quantization provides minimum performance reduction (0.1%). 

Our work opens the scope for running radiologic image processing models on devices like mobile-phones and embedded boards with comparable accuracy.

Low-Powered Deep Learning Models for Radiology Imaging

The new EuSoMII Annual Meeting 2020 is a very nice live format with hosts and moderators in live video streaming  plus speakers connected remotely from Europe and US. The Program includes a morning session with Didactic Lectures and the possibility to do an examination at the end. A great afternoon session will follow: 2 great European speakers  and a Joint Session with our US partner  SIIM (Society for Imaging Informatics in Medicine). [Visit the Event Page](https://www.eusomii.org/events/eusomii-annual-meeting-2020/https://www.eusomii.org/events/eusomii-annual-meeting-2020/#:~:text=The%20new%20EuSoMII%20Annual%20Meeting,an%20examination%20at%20the%20end.)

[Introduction of the speakers](https://drive.google.com/file/d/12c7ObMM72HgA6TYZ1j8NWTqCcHumXurG/view?usp=sharing)

Sponsors:
[SECTRA PLATINUM EXCLUSIVE](https://medical.sectra.com/solutionarea/ai-medical-imaging/) | [QUIBIM Platinum](https://quibim.com/) | [PHILIPS Platinum](https://www.philips.com/global) | [ROBOVISION Platinum](https://robovision.ai/) | [BRACCO Imaging GOLD](https://www.braccoimaging.com/) | [Contextflow Silver](https://contextflow.com/) | [TeraRecon Silver](https://www.terarecon.com/) | [InferVISION](https://global.infervision.com/)