At the IEEE TENCON2019 held from October 17th 2019 to October 20th 2019 at Cochin, India, I was privileged to present our work on Lung Nodule Detection using the Intel® Distribution of OpenVINO™ toolkit. TENCON 2019 had approximately 500 papers in 10 regular tracks and 25 Special Session with most papers focusing on AI, with 4 days packed with tutorial sessions, paper presentation and keynote talks. Our paper was presented under the AI for healthcare session. TENCON is a premier international technical conference of IEEE Region 10. The theme for this year’s conference was Technology, Knowledge and Society. And our work perfectly blends with this year’s theme. After the presentation, the stage was thrown open to Q & A session with questions related to data the preprocessing stage and training phase.
With the advancement of AI in the field of medical imaging, medical diagnosis is getting faster and viable for medical practitioners especially for cancer diagnosis. Earlier deep learning (DL) solutions had to be deployed on high performance computing hardware accelerators for achieving real-time performance. But by utilizing optimizations on Intel Core™ and Intel Xeon® processors with the Intel Distribution of OpenVINO toolkit (Open Visual Inference and Neural Network Optimization), it is possible to deploy DL models with accelerated performance in comparison with running Tensorflow or Caffe models on CPU processors. In TENCON 2019, we presented our work demonstrating how we ported QuEST’s DetectNet based Deep Learning Model with a hardware accelerator-specific custom layer for Lung nodule detection trained on LIDC dataset using Intel Distribution of OpenVINO toolkit, and deployed the same model on Intel Core i7 and Intel Xeon processors. The DL model we deployed had high sensitivity and precision with only low false positive rate. We achieved a 33x performance improvement by running the Intel Distribution of OpenVINO toolkit model on an Intel Core i7-8700 processor in comparison with an un-optimized baseline model. With our model, each chest Computerized Tomography (CT) scan image slice sized 1024x1024x3 took roughly 0.2304 seconds for execution on an Intel Core i7-8700 processor. These results show the advantage of deploying AI solutions on Intel-based processors especially in hospitals and healthcare institutions where Intel Architecture is very prevalent. Please refer to the below table for details regarding HW configuration used in the experiments.