Deep learning has become a powerful tool in solving image classification problems1 and has been widely used in medical image analysis.2 Convolutional neural networks (CNNs) show the potential for automated classification of different cancer lesions, such as lung cancer,3 skin cancer,4 and breast cancer.5 But the lack of interpretability and explainability makes CNNs less understandable.6 To push the envelope of CNN-based medical diagnostic methods, more rigorous neural network models and training methods must be developed to provide both visually interpretable and explainable analysis.7–9
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.
Oral cancer is a common disease, particularly in low- and middle-income countries. Early detection of oral cancer is believed to be the most effective way to prevent it. Automatic oral cancer image classification algorithms based on machine learning enable a system to learn from previous data and, based on the learning, predict and give results on new unseen data. However, oral cancer datasets captured from high-risk populations are often unbalanced since there are many more normal cases than benign, premalignant, and malignant cases. Due to the influence of the larger majority classes featured on machine learning training framework, the classifier results may be biased toward the over-represented class (or classes). The discrimination of minority classes is also of great clinical significance. For example, a false positive result for benign lesions in cancer screening will result in unnecessary psychological stress, medical procedures to patients, and increased clinical workloads. Similarly, classifiers also need high sensitivity because they are designed for cancer screening. Therefore, one of the big challenges is handling class imbalance, especially for multiple category classification.
Oral cancer incidence and death rates are an ongoing concern in low- and middle-income countries (LMICs). LMICs suffer the majority of worldwide new cancer cases and deaths along with five-year survival rates in some locations.1,2 The major risk factor for oral squamous cell carcinoma and oral submucous fibrosis in many LMICs is betel quid chewing, with or without the inclusion of tobacco.3,4 Furthermore, oral and oropharyngeal cancers are increasing in younger populations (especially men) in middle- to high-income communities (MHICs) due to the spread of human papillomavirus (HPV), mostly affecting the tonsils and base of the tongue (BOT).5–7
Convolutional neural networks have demonstrated excellent performance in oral cancer detection and classification. However, the end-to-end learning strategy makes CNNs hard to interpret, and it can be challenging to fully understand the decision-making procedure. Additionally, reliability is also a significant challenge for CNN based approaches. In this study, we proposed a neural network called the attention branch network (ABN), which combines the visual explanation and attention mechanisms to improve the recognition performance and interpret the decision-making simultaneously. We also embedded expert knowledge into the network by having human experts manually edit the attention maps for the attention mechanism. Our experiments have shown that ABN performs better than the original baseline network. By introducing the Squeeze-and-Excitation (SE) blocks to the network, the cross-validation accuracy increased further. Furthermore, we observed that some previously misclassified cases were correctly recognized after updating by manually editing the attention maps. The cross-validation accuracy increased from 0.846 to 0.875 with the ABN (Resnet18 as baseline), 0.877 with SE-ABN, and 0.903 after embedding expert knowledge. The proposed method provides an accurate, interpretable, and reliable oral cancer computer-aided diagnosis system through visual explanation, attention mechanisms, and expert knowledge embedding.
Oral cancer is one of the leading causes of cancer-related deaths, especially in South Central Asia and Melanesia, accounting for 377,713 new cases and 177,757 new deaths in 2020, according to GLOBOCAN.1 It is highly prevalent in India, where the incidence rate in males was 13.9 per 100,000 and the mortality rate in females was 7.7 per 100,000 in 2018.2 In middle- and low-income countries such as India, the five-year survival rate is due to late diagnosis, according to the paper published in 2014.3 Therefore, point-of-care screening platforms and algorithms are in great need.