Machine Learning Replaces Conventional Processing of Raw Optical Coherence Tomography Spectrometer Data

Abstract

Purpose : Conventional processing of OCT spectrometer data is comprised of isolation of the fringe signal, resampling from frequency to wave space, dispersion compensation, and Fourier transformation. Resulting A and B-scans are fraught with speckle noise, requiring further processing to become clinically usable. We propose using machine learning (ML) digital signal processing (DSP) as an alternative direct pathway to produce OCT B-scans.

Methods : We applied a 1D neural network autoencoder built with the Keras Python library running on top of TensorFlow to process raw spectrometer data into single A-scans. Two ML algorithms were trained using entirely unprocessed spectra; one to post-processed and the second to averaged A-scans. The training A-scans were a combination of spectra obtained from a layered phantom, and spectra from a retinal imagery (Bioptigen OCT). In total, 15 different image volumes were used. Four were reserved for testing and validation. The 11 training volumes provided 1.5 million spectra lines and their corresponding processed counterparts for training. From this, 750,000 were randomly selected on each epoch for training. Equal distribution across all volumes reduced the probability of overfitting.

Results : Assessed subjectively, when trained to the averaged dataset, ML-DSP produced B-scans directly from raw spectrometer data with less speckle noise (Figure 1: Phantom, Figure 2: Retina). A-scan to A-scan alignment within the assembled B-Scan is reduced in the phantom image, and poor in the retinal image.

Conclusions : We have demonstrated that ML-DSP is capable of processing raw spectrometer data directly to produce B-scans. Future work will address A-scan alignment by expanding to a 2D ML-DSP engine, processing entire B-scans instead of individual A-scans. Future ML engines may be enhanced to provide measurements, segmentation, angiography, Doppler, and even diagnosis directly from raw spectral data; bypassing numerous conventional processing steps. The successful implementation of this technique could be a substantial advance in OCT technology.

The opinions or assertions contained herein are the private ones of the authors/speaker and are not to be construed as official or reflecting the views of the Department of Defense, the Uniformed Services University of the Health Sciences or any other agency of the U.S. Government. Patent Pending

This is a 2020 ARVO Annual Meeting abstract.