Rapid and Label-Free Histopathology of Oral Lesions Using Deep Learning Applied to Optical and Infrared Spectroscopic Imaging Data

Abstract

Oral potentially malignant disorders (OPMDs) are precursors to over 80% of oral cancers. Hematoxylin and eosin (H&E) staining, followed by pathologist interpretation of tissue and cellular morphology, is the current gold standard for diagnosis. However, this method is qualitative, can result in errors during the multi-step diagnostic process, and results may have significant inter-observer variability. Chemical imaging (CI) offers a promising alternative, wherein label-free imaging is used to record both the morphology and the composition of tissue and artificial intelligence (AI) is used to objectively assign histologic information. Here, we employ quantum cascade laser (QCL)-based discrete frequency infrared (DFIR) chemical imaging to record data from oral tissues. In this proof-of-concept study, we focused on achieving tissue segmentation into three classes (connective tissue, dysplastic epithelium, and normal epithelium) using a convolutional neural network (CNN) applied to three bands of label-free DFIR data with paired darkfield visible imaging. Using pathologist-annotated H&E images as the ground truth, we demonstrate results that are 94.5% accurate with the ground truth using combined information from IR and darkfield microscopy in a deep learning framework. This chemical-imaging-based workflow for OPMD classification has the potential to enhance the efficiency and accuracy of clinical oral precancer diagnosis.

Keywords: deep learning; discrete frequency infrared microscopy; multimodal imaging; oral potentially malignant lesions; precancerous condition.

Figure 1

A combined IR and darkfield microscopy workflow for histopathology of oral potentially malignant tissues. The workflow combines the IR and DF images with pathologist annotations for the training dataset. Patches of images are passed to the classifier model to generate the final segmented image.

Figure 2

Discrete frequency IR imaging absorbance variation across oral tissue. (A) Normalized intensity distribution with outliers removed for 3 classes for absorbance at 1238 cm⁻ (left) and 1546 cm⁻¹ (right). Representative whole biopsy IR images at (B) 1238 cm⁻¹ and (C) 1546 cm⁻¹. (D) Darkfield visible image of whole unstained biopsy section. (E) H&E-stained image of section adjacent to darkfield and IR imaged section. (F) class annotations for reference. Scale bar: 500 µm.

Figure 3

Comparison of test dataset confusion matrices, in percentage, for models trained on different imaging techniques. Confusion matrix for a model trained (A) exclusively using dark field images, (B) solely using IR images, and (C) by combining both IR and dark field images, showcasing the potential synergy between the two imaging methods.

Figure 4

Comparative visualization of the deep learning model’s accuracy in identifying dysplasia. (A) Dysplastic sample; (B) non-dysplastic sample. Each set, moving from left to right, includes: an IR image, a dark field visible image, the ground truth annotation, the model’s prediction, and the adjacent H&E image.