NVIDIA and the National Institutes of Health (NIH) joined forces last year to create clinical deep learning tools in diagnostic and interventional imaging, mostly focusing on prostate cancer, leveraging imaging and pathology data to better understand the physical and the cellular structure of the disease.

NIH radiologists gathered 465 mpMRI data from various medical centers, representing multiple MRI vendors including Siemens, Philips, GE and many other center-specific MRI protocols.

A radiologist experienced in prostate MRI was able to manually trace the prostate boundaries in three planes on T2-weighted MRI. NVIDIA’s scientists then used the Clara Train SDK to foster a 3D deep learning-based pipeline based on a previous workflow. They used a hybrid 2D-3D neural network, evaluating 98 patients of unseen data. This process attained a 0.922 DICE score, as opposed to the DICE score between different radiologists’ annotations, which was 0.919. Therefore, this approach obtained a similar performance to the highly experienced radiologists that annotated the data.

Data provided by NIH trained the model that was locked and evaluated on multiple prostate MRI datasets from public challenges. The model gained similar or better functionality compared to the state-of-the-art algorithms on every single dataset. It achieved a DICE score of 0.894 versus 0.904 of state-of-the-art algorithms on average.

NVIDIA and NIH incorporated and directed the procedure using NVIDIA V100 GPUs with the TensorFlow.

The joint effort has yielded a tool that can potentially help improve diagnostic and treatment accuracy for prostate cancer patients.