Automated Histopathological Analysis
One major challenge of deep learning in medical imaging is that large numbers of images are not readily available. The problem becomes easier in digital histopathology. After digitalization with high resolution, a whole image slide typically contains billions of pixels from which hundreds of thousands of image patches can be extracted. Machine learning /deep learning thus becomes a promising tool to reduce pathologists’ workload and facilitate personalized medicine.
For example, examination of blood smear images is a routine work in every hospital, while the manual analysis is time-consuming and tedious. To address the issue, we are building a machine learning tool to detect and classify blood cells in smear images. Further, we combine the morphological features and corresponding optical scattering data of blood cells to address specific clinical conditions, such as leukemia and diabetes. It is expected that the histopathological analysis can be streamlined via the developed machine learning solution.
Detecting Cervical Spine Fracture in CT
In cases of head/neck trauma (automobile accidents, falls among the elderly, etc.), patients must be evaluated for potential cervical spine fracture due to the serious complications that can occur in an untreated injury. The most common screening tool is non-contrast CT. In cases where a fracture is detected, CT Angiography is ordered to assess the presence of arterial injury.
With an aging population, the number of c-spine scans is rapidly growing. While the relative frequency of fractures is low (<10%), radiologists must rapidly evaluate every scan, increasing the clinical burden. At some centers, these studies are evaluated at the scanner, forcing the radiologist to leave the reading room and interrupting their workflow. In this work, we aim to develop a tool to detect the presence of c-spine fracture in CT to lessen this burden, reduce visits to the scanner, and reprioritize the worklist.
Automated Body Composition From CT Scans
Body composition, i.e. the proportion of muscle and fat within the body, is an important biomarker with a range of potential applications medicine, such as monitoring cancer progression and surgical planning.
CT scans are already routinely captured in these situations and contain sufficient information to distinguish muscle and fat from other tissues, but doing so currently requires tedious manual or semi-automated analysis by clinicians. In this work, we are leveraging deep learning models to develop a fully-automated tool that selects the relevant slices of a CT scan and delineates regions of muscle and fat in an efficient and repeatable way.