In a recent study presented in March at the Strata Data Conference in San Jose, California, Dell EMC researchers Mauro Damo, Wei Lin, Ronaldo Braga, and William Schneider demonstrated that using machines to analyze medical images—specifically diseased human organs—can minimize medical errors and speed up disease diagnosis.
Using more than 5,000 magnetic resonance images (MRI) scanned along the transverse plane from the Cancer Imaging Archive, the Dell EMC data scientists attempted to identify bladder cancer using a six-layer convolution neural network (CNN). Such networks have successfully identified faces, objects, and traffic signs, and are used to power vision in robots and self-driving cars.
The Dell EMC study tracked four out of fourteen types of primary tumors—T2a, T2b, T3a, and T4a—in patients already diagnosed with bladder cancer. Algorithms were also used to identify significant differences among the images to assess what features could be relevant for bladder cancer detection. Bladder cancer, the fourth most common cancer in men, impacts 430,000 people and their families each year around the world. The direct medical cost of bladder cancer care was $125B in 2010 globally.
To conduct the research, team members used powerful hardware and open-source software, including:
- Dell EMC PowerEdge servers with Intel Xeon E5-2680 processors @ 2.7GHz with 8 cores and 384 GB
- NVIDIA GRID K2 with 2 GPUs
- Python open-source script language
- TensorFlow 1.4 open-source deep learning package
- SimpleITK open-source image transformation package
Researchers learned that GPU and CPU memory are more relevant than cycles in hardware, and that out of memory errors are very common when working with medical data. For models that take weeks to run, it’s better to improve memory and cycles or use a distributed platform; otherwise, it’s preferable to have more memory to fit all your weights initialization and mini batch processes.
Using the six-layer convolutional neural network (CNN), the study’s deep learning model improved Top 1 accuracy from 72.3 percent to 81.3 percent, showing the potential to explore similar techniques in practical applications. Researchers noted that more data in the early stages of cancer, the analysis of more types of primary tumors, and the use of independent CNN models for all planes (coronal, transverse, and sagittal) would likely improve results.
To learn more about how deep learning can enhance research and processes in healthcare, financial services, retail, manufacturing, and other industries, explore Dell EMC Ready Solutions for AI, Machine and Deep Learning or contact your Dell EMC representative at 1-866-438-3622. We’d be glad to show you a demo in our Dell EMC HPC & AI Innovation Lab, which is at the cutting edge of machine learning, testing new technologies, and tuning algorithms and applications to keep pace with the constantly evolving landscape.