The use of optical coherence tomography and convolutional neural networks to distinguish normal and abnormal oral mucosa

Abstract

Incomplete surgical resection of head and neck squamous cell carcinoma (HNSCC) is the most common cause of local HNSCC recurrence. Currently, surgeons rely on preoperative imaging, direct visualization, palpation and frozen section to determine the extent of tissue resection. It has been demonstrated that optical coherence tomography (OCT), a minimally invasive, nonionizing near infrared mesoscopic imaging modality can resolve subsurface differences between normal and abnormal head and neck mucosa. Previous work has utilized two-dimensional OCT imaging which is limited to the evaluation of small regions of interest generated frame by frame. OCT technology is capable of performing rapid volumetric imaging, but the capacity and expertise to analyze this massive amount of image data is lacking. In this study, we evaluate the ability of a retrained convolutional neural network to classify three-dimensional OCT images of head and neck mucosa to differentiate normal and abnormal tissues with sensitivity and specificity of 100% and 70%, respectively. This method has the potential to serve as a real-time analytic tool in the assessment of surgical margins.

Keywords: head and neck neoplasms; margins of excision; optical coherence tomography; oral cancer; squamous cell carcinoma; squamous cell carcinoma of head and neck.

Figure 1:

Optical and electrical schematic for the VCSEL-SS OCT system and scanning probe utilized for this study. ODL: Optical delay line used to match the optical path length of the sample arm, FC: In line fiber optic coupler used to split and combine the laser light used in the interferometer, D: Balanced photodiode used to detect the interference OCT signal, C1,2: Fiber optic in-line circulator used to direct the representative sample and reference beams.

Figure 2:

Representative areas imaged for two of the six HNSCC cases. Green bars and arrows indicate scanned area and scanning direction. (a-c) Series of 3-D OCT volumes acquired from anterior to posterior aspect of the resected tongue specimen. (d-f) Series of 3-D OCT volumes acquired for the superior and anterior aspect of the resected tonsil and soft palate specimen.

Figure 3:

(a,d) Visible light photograph of a resected specimen with red bars and arrows indicating scanned area and scanning direction. (b,e) Corresponding H&E histology sections. (c,f) Corresponding false colored OCT image that has been preprocessed for convolutional neural network training.

Figure 4:

(a) Image of the normalized OCT power spectrum data of a single B-scan. (b) Corresponding histogram of the normalized power spectral data. (c) Power spectrum data of the representative B-scan with rescaled colormap.

Figure 5:

(a) Schematic block diagram of AlexNet showing convolution, max pooling and fully connected layers of the CNN. (b) 96 convolutional 11 x 11 x 3 kernel filters. Adapted from “ImageNet Classification with Deep Convolutional Neural Networks” by Krizhevsky A. et al. (2012).

Figure 6:

Accuracy and loss training record for the supervised transfer learning of AlexNet with the OCT head and neck images obtained in this study.

Figure 7:

CNN classification probability output and false color mapping

Figure 8:

(a-c): Labeled and orientated visible light images of a tongue specimen scanned with 3-D OCT. Green bars and arrows indicate scanned area and scanning direction. (d-f) Corresponding H&E stained histology sections (g-i) CNN classification of the scanned area indicated in (a), (b) and (c), with the Z axis as the classified probability and the X axis as the B-Scan number out of 1000 total B-scans in a single OCT-3-D volumetric data acquisition. The Y axis was arbitrary determined for graphical visualization.

Figure 9:

Spatial representation of neural network classification for 3-D OCT volumes.

Equation 1:

Calculated probability for labeled OCT 3-D volumes as normal or abnormal