(208a) Discrete Frequency Infrared Chemical Imaging for Histochemical Interpretation of Renal Biopsy

Current histopathological methods rely largely on manual optical inspection of stained tissue providing only extremely limited chemical information about the system. Chemical imaging techniques such as Fourier transform infrared (FTIR) imaging have been used to overcome this limitation. Although FTIR imaging does allow for chemical information to be gathered without the use of stains, the time and data size requirements to do so at the necessary spectral and spatial resolution are high. Quantum cascade laser (QCL) based discrete frequency infrared (DFIR) imaging overcomes many of these limitations by acquiring absorption at only the wavenumbers of interest. Kidney diseases lead to an estimated 5-10 million deaths annually with many of these caused by glomerular diseases. Diagnosis of these diseases is hampered by overlapping morphological characteristics in hematoxylin and eosin (H&E) stained images, therefore more expensive techniques may be required. In this work a combination of DFIR and darkfield visible imaging and deep learning (DL) have been used to first identify glomeruli then to classify the glomeruli into 5 different disease states: (1) normal (2) diabetic nephropathy (3) light chain amyloidosis (4) light chain deposition disease (5) fibrillary glomerulonephritis. A deep convolutional U-net was used for glomeruli classification. Successful classification of glomeruli and H&E stainless staining were demonstrated using only a darkfield image combined with 3 discrete wavenumber images. Disease state classification was performed using full QCL based IR point spectra acquired within the DL identified glomeruli. The combination visible and DFIR DL methodology conforms with a clinically relevant timescale.