Automated monitoring and screening of various physiological signals is an indispensable tool in modern medicine. However, despite the preponderance of long-term monitoring and screening modalities for certain vital signals, there are a significant number of applications for which no automated monitoring or screening is available. For example, patients in need of urinary catheterization are at significant risk of urinary tract infections, but long-term monitoring for a developing infection while a urinary catheter is in place typically requires a caregiver to frequently collect urine samples which then must be transported to a laboratory facility to be tested for a developing infection. Disruptive technologies at the intersection of lens-free imaging, fluidics, image processing, computer vision and machine learning offer a tremendous opportunity to develop new devices that can be connected to a urinary catheter to automatically monitor urinary tract infections. However, novel image reconstruction, object detection and classification, and deep learning algorithms are needed to deal with challenges such as low image resolution, limited labeled data, and heterogeneity of the abnormalities to be detected in urine samples. This project brings together a multidisciplinary team of computer scientists, engineers and clinicians to design, develop and test a system that integrates lens-free imaging, fluidics, image processing, computer vision and machine learning to automatically monitor urinary tract infections. The system will take a urine sample as an input, image the sample with a lens-free microscope as it flows through a fluidic channel, reconstruct the images using advanced holographic reconstruction algorithms, and detect and classify abnormalities, e.g., white blood cells, using advanced computer vision and machine learning algorithms. Specifically, this project will: (1) design fluidic and optical hardware to appropriately sample urine from patient lines, flow the sample through the lens-free imager, and capture holograms of the sample; (2) develop holographic image reconstruction algorithms based on deep network architectures constrained by the physics of light diffraction to produce high quality images of the specimen from the lens-free holograms; (3) develop deep learning algorithms requiring a minimal level of manual supervision to detect various abnormalities in the fluid sample that might be indicative of a developing infection (e.g., the presence of white bloods cells or bacteria); and (4) integrate the above hardware and software developments into a system to be validated on urine samples obtained from patient discards against standard urine monitoring and screening methods.
This project could lead to the development of a low-cost device for automated screening and monitoring of urinary tract infections (the most common hospital and nursing home acquired infection), and such a device could eliminate the need for patients or caregivers to manually collect urine samples and transport them to a laboratory facility for testing and enable automated long-term monitoring and screening for UTIs. Early detection of developing UTIs could allow caregivers to preemptively remove the catheter before the UTI progressed to the point of requiring antibiotic treatment, thus reducing overall antibiotic usage. The technology to be developed in this project could also be used for screening abnormalities in other fluids, such as central spinal fluid, and the methods to detect and classify large numbers of cells in an image could lead to advances in large scale multi-object detection and tracking for other computer vision applications.