Researchers at Purdue University have developed AI-driven technology that utilizes smartphone cameras to detect and diagnose medical conditions like anemia more quickly and accurately than specialized medical equipment. This innovation aims to leverage the numerous sensors present in smartphones for remote healthcare access. The lead researcher, Young Kim, explained that the goal is to capitalize on the at least 15 sensors in smartphones to provide healthcare solutions beyond traditional clinical settings.
The technology involves employing a smartphone camera to perform hyperspectral imaging, which captures all visible light wavelengths in each pixel. This method has potential applications in detecting various skin and retinal conditions, as well as certain types of cancer. Although hyperspectral imaging has been explored for healthcare, it often involves bulky, slow, and expensive specialized equipment.
The Purdue team combined deep learning and statistical techniques with knowledge of light-tissue interactions to develop an algorithm that can reconstruct the full spectrum of visible light for each pixel in smartphone images. This patent-pending approach allows for rapid hyperspectral imaging using a common smartphone camera.
In their study, the researchers tested the smartphone-based hyperspectral imaging against commercially available specialized equipment. They gathered data about blood oxygen movement in volunteers’ eyelids, human tissue models, and chick embryos. The smartphone camera provided hyperspectral information more quickly, cost-effectively, and with comparable accuracy to the specialized equipment, achieving results in milliseconds that would take conventional hyperspectral imaging minutes to produce.
The process involves capturing ultra-slow-motion video using a smartphone camera, processing it through a machine learning algorithm, and inferring full-spectrum information for each pixel. This information is then used to measure hemodynamic parameters like oxygenated and deoxygenated hemoglobin. The algorithm is trained using both smartphone images and corresponding hyperspectral images, making use of equations derived from tissue optics. This «informed learning» approach requires a smaller training dataset than conventional machine learning methods.
Unlike traditional hyperspectral imaging equipment, which must gather extensive data and often compromise either spectral or temporal resolution, the Purdue team’s method begins with much smaller video files. This allows for high spatial and spectral resolution simultaneously.
The researchers are actively exploring applications of this technology in various mobile health scenarios, such as cervix colposcopy and retinal fundus imaging. The innovation has the potential to revolutionize healthcare access by making diagnostic capabilities available through widely accessible smartphone technology.
Article written by Michael Barbela