Validation of a Point-of-Care Optical Coherence Tomography Device with Machine Learning Algorithm for Detection of Oral Potentially Malignant and Malignant Lesions

Abstract

Non-invasive strategies that can identify oral malignant and dysplastic oral potentially-malignant lesions (OPML) are necessary in cancer screening and long-term surveillance. Optical coherence tomography (OCT) can be a rapid, real time and non-invasive imaging method for frequent patient surveillance. Here, we report the validation of a portable, robust OCT device in 232 patients (lesions: 347) in different clinical settings. The device deployed with algorithm-based automated diagnosis, showed efficacy in delineation of oral benign and normal (n = 151), OPML (n = 121), and malignant lesions (n = 75) in community and tertiary care settings. This study showed that OCT images analyzed by automated image processing algorithm could distinguish the dysplastic-OPML and malignant lesions with a sensitivity of 95% and 93%, respectively. Furthermore, we explored the ability of multiple (n = 14) artificial neural network (ANN) based feature extraction techniques for delineation high grade-OPML (moderate/severe dysplasia). The support vector machine (SVM) model built over ANN, delineated high-grade dysplasia with sensitivity of 83%, which in turn, can be employed to triage patients for tertiary care. The study provides evidence towards the utility of the robust and low-cost OCT instrument as a point-of-care device in resource-constrained settings and the potential clinical application of device in screening and surveillance of oral cancer.

Keywords: artificial neural network; optical coherence tomography; oral cancer; oral potentially malignant lesions; oral squamous cell carcinoma; pre-malignant lesions.

Figure 1

Study design. The subjects were: (A) recruited from low resource settings (oral cancer screening camps), dental hospitals and tertiary cancer center. (B) A portable Optical Coherence Tomography (OCT) system, (C)was used to capture oral mucosal lesion images, after oral physician consultation. The OCT images were recorded in laptop and used for image pre-processing and automated image analysis (D,E). The subjects underwent incision/excision biopsy (if indicated) for histopathological diagnosis (F,G). The OCT images were then analyzed by automated image processing and algorithm/artificial intelligence (H) based classification and compared with histological diagnosis.

Figure 2

ANN-SVM based image analysis pipeline. Image data sets selected for image pre-processing and segmentation (A) after quality evaluation. The image was enhanced by Gaussian filter. Template images were used to segment region of interest of upper and lower sections by 2D-cross correlation. Feature extraction (B) was performed on segmented images by multiple artificial neural networks (ANN). The extracted features were used for developing Support Vector Machine model (SVM) model for each ANN feature vectors (C). The receiver operating characteristic (ROC) curve analysis was performed in cross-validation data set to find out optimal cut-off score for classification. The SVM-model validated in test data set and classified according cut-off score. The ANN-SVM models were developed for stepwise classification- initially for classifying OSCC from dysplastic/normal/benign lesions and then dysplastic from benign/normal lesions. OSCC: Oral squamous cell carcinoma, HGD: High grade dysplasia–Moderate/Severe dysplasia, LGD: Low grade dysplasia- Mild dysplasia, hyperplasia, ANN: Artificial neural network, 2D- 2 dimensional, SVM: Support vector machine, ROC: Receiver Operating Characteristic.

Figure 3

Clinical, OCT & histology images. Clinical (A) and OCT (B) images were captured for all the subjects and the biopsy tissues collected (wherever indicated) were assessed by histopathology (C). Histology images were taken at 100× resolution (scale bar = 100 µm) using Nikon DSFi2 and NIS elements D4 20.0. The non-dysplastic lesions shown were histologically diagnosed with lichen planus, pyogenic granuloma, and hyperkeratosis. Normal buccal mucosa images were taken from healthy volunteer without any habit history. Representative images of all dysplastic grades and a buccal oral squamous cell carcinoma (OSCC) are also depicted.

Figure 4

OCT algorithm prediction score distribution between the different patient cohorts. Box and whisker plot depicting the algorithm scores of oral squamous cell carcinoma (OSCC), dysplasia and normal/benign lesions. The score significantly increases (p < 0.005) as the disease progresses. Graph shows median and inter-quartile range.

Figure 5

Accuracy of the various ANN-SVM models in delineating the patient cohorts. The training and test accuracy of the 14 neural networks used in the study in delineating cancer from dysplasia/non-dysplastic lesion and dysplasia from non-dysplastic lesions were depicted. The size of circle represents the size of neural network. The less overfitting models were NASNetMobile and DenseNet-201.

Figure 6

Sensitivity and specificity of neural networks. DenseNet-201 and NASNetMobile showed best sensitivity/specificity in delineating cancer vs others whereas DenseNet-201 was best in differentiating dysplasia vs non-dysplasia lesions.

Figure 7

Decision tree for incisional biopsy from OCT diagnostic system. The decision tree consists of Test 1, which will identify OSCC patients. Test 2 and Test 3 can be used to identify patients with dysplasia and high-grade dysplasia (HGD) respectively. The combined sensitivity of these three tests ranges between 93–95% for delineating oral cancer and dysplasia. This decision tree can be used to triage high-risk patients in oral cancer screening camps. ANN: Artificial neural network.