Medical imaging technologies, such as MRI and CT scanners, have benefited from the recent breakthroughs in Artificial Intelligence (AI) to produce high-quality images faster than ever before. Traditionally, these technologies need to take samples of an object from many different views in order to reconstruct the original 3D object with a high resolution (Figure 1). Generally, the more views that are sampled, the higher quality the final image reconstruction. This presents the classic tradeoff between time and quality—you can do the job quickly or with high resolution— but not both.
Compressed Sensing (CS) bends this tradeoff by introducing some changes to the traditional approach. First, CS assumes that the image is “sparse” (or at least somewhat sparse). In general, most images have some level of sparsity—that is, there is some degree of order or structure in the image. Figure 2 demonstrates that an image with random noise is not sparse (left), but a painting has some sparsity (right). For example, the hat and dress of the young girl and the walking path she travels have structure to them. In the early 21st century, compressed sensing researchers proved that for sparse images such as these, the original image could be reconstructed with high quality if the signal were sampled incoherently instead of the traditional method of regular sampling (white dots in Figure 3). Critically, the number of samples (white dots) needed for the compressed sensing approach is substantially lower than in the traditional sampling approach (Figure 3, left versus right image).
CS techniques have been integrated into clinical CT and MR scanners over the last several years. More recently, deep learning AI algorithms have been combined with CS to improve the quality of the final reconstructed image. Figure 4 shows a combined CS/AI approach called W-Net that yields a good quality MRI and is five times faster than the traditional approach (Souza and Frayne, 2018).
Intel plays a vital role in the adoption of AI into the medical imaging industry. Intel CPUs are a key component in most of the medical imaging scanners today. We work directly with healthcare companies to make sure their workloads run optimally on Intel hardware. With the emergence of AI, Intel has made great strides in bringing cutting edge techniques like CS/AI to medical imaging devices.
In fact, Intel recently partnered with Philips Healthcare to optimize a Philips AI algorithm for CS that resulted in a 54x speedup on the same Intel CPU by using the Intel® Distribution of the OpenVINO toolkit (Figure 5). We invite you to read about the Intel-Philips collaboration and learn more about how Compressed Sensing and Artificial Intelligence can allow medical scanners to perform faster high-quality images.
- Souza R and Frayne R. “A Hybrid Frequency-domain/Image-domain Deep Network for Magnetic Resonance Image Reconstruction”, arXiv 1810.12473, 2018.
- Pezzotti N, de Weerdt E, Yousefi S, Elmahdy MS, van Gemert J, Schülke C, Doneva M, Nielsen T, and Kastryulin S, Lelieveldt BPF, van Osch MJP, and Staring M. “Adaptive-CS-Net: FastMRI with Adaptive Intelligence”, arXiv 1912.12259, 2019.
- Reina GA, Stassen M, Pezzotti N, Moolenaar D, Kurtaev D, Khowala A. “Philips Healthcare Uses the Intel® Distribution of OpenVINO™ Toolkit and the Intel® DevCloud for the Edge to Accelerate Compressed Sensing Image Reconstruction Algorithms for MRI”, 2020.