Accurate histopathology from low signal-to-noise ratio spectroscopic imaging data

Fourier Transform Infrared (FT-IR) spectroscopic imaging is emerging as an automated alternative to human examination in studying development and disease in tissue. The technology's speed and accuracy, however, are limited by the trade-off with signal-to-noise ratio (SNR). Signal processing approaches to reduce noise have been suggested but often involve manual decisions, compromising the automation benefits of using spectroscopic imaging for tissue analysis. In this manuscript, we describe an approach that utilizes the spatial information in the data set to select parameters for noise reduction without human input. Specifically, we expand on the Minimum Noise Fraction (MNF) approach in which data are forward transformed, eigenimages that correspond mostly to signal selected and used in inverse transformation. Our unsupervised eigenimage selection method consists of matching spatial features in eigenimages with a low-noise gold standard derived from the data. An order of magnitude reduction in noise is demonstrated using this approach. We apply the approach to automating breast tissue histology, in which accuracy in classification of tissue into different cell types is shown to strongly depend on the SNR of data. A high classification accuracy was recovered with acquired data that was similar to 10-fold lower SNR. The results imply that a reduction of almost two orders of magnitude in acquisition time is routinely possible for automated tissue classifications by using post-acquisition noise reduction.