We developed an artificial intelligence model to assist physicians in the detection of nasopharyngeal malignancies and provide biopsy guidance by applying deep learning to nasopharyngeal endoscopy examination. We demonstrated that eNPM-DM was superior to oncological experts in detecting nasopharyngeal malignancies. Of note, eNPM-DM exhibited encouraging performance in nasopharyngeal malignancy segmentation.
Currently, the emerging field of machine learning, especially deep learning, has exerted significant impact on medical imaging. In general, deep learning algorithms recognize important features of images and properly giving weight to these features by modulating its inner parameters to make predictions for new data, thus accomplishing identification, classification or grading [21] and demonstrating strong processing ability and intact information retention [22], which is superior to previous machine learning methods [17]. Superiority of computer-aided diagnosis based on deep learning have recently been reported for a wide spectrum of diseases, including gastric cancer [23], diabetic retinopathy [16], cardiac arrhythmia [24], skin cancer [17] and colorectal polyp [25]. Notably, a wide variety of image types were explored in these studies, i.e., pathological slides [19, 23], electrocardiograms [24], radiological images [18, 26] and general pictures [17]. The deep learning method exhibited outstanding performance in most of the competitions between artificial intelligence and experts even though these medical images were captured by various types of equipment and presented in different forms, suggesting enormous potential of deep learning in auxiliary diagnoses. A well-trained algorithm for a specific disease can increase the accuracy of diagnosis and working efficiency of physicians, liberating them from repetitive tasks.
Recently, deep learning has been extensively used in the differential diagnosis of gastrointestinal disease in endoscopic images. Tomohiro et al. developed a convolutional neural network for detecting gastric cancer [27] and Helicobacter pylori infection based on endoscopic images [28]. Moreover, an artificial intelligence model was trained on endoscopic videos to differentiate diminutive adenomas from hyperplastic polyps, thus realizing real-time differential diagnosis [29]. Given that endoscopic examination is indispensable for biopsy and important for decision making in a clinical setting, developing tools for endoscopic auxiliary diagnosis can dramatically increase physicians’ working efficiency via rapid recognition and biopsy guidance, especially in patients with multi-lesions or mixed lesions [30]. Given the illusive mass caused by adenoid/lymphoid hyperplasia, it is desirable to recognize nasopharyngeal malignancies using artificial intelligence tools. However, limited studies on deep learning methods in nasopharyngeal disease differentiation have been performed based on endoscopic images to date [31]. To this end, this study has taken advantage of the abundant resource of nasopharyngeal endoscopic images at our centre and the advanced methods to develop the targeted model.
Endoscopic examination is particularly indispensable for biopsy in participants at risk of nasopharyngeal malignancies. However, currently, no additional approaches are applied to the screening of nasopharyngeal malignancies except EBV serological test [32]. Our eNPM-DM outperformed experts in distinguishing nasopharyngeal malignancies from benign diseases using far less time with encouraging sensitivity and specificity. eNPM-DM was trained and fine-tuned on numerous images that covered patients diagnosed at our centre over 8 years and exhibited encouraging performance in a shorter learning period, suggesting that eNPM-DM ‘learned’ efficiently and was highly productive.
Over diagnosis is the major cause of misdiagnosis for both eNPM-DM and oncologists, suggesting that the model might learn object recognition in the same manner as a human. For example, both could distinguish different objects based on the texture, roughness, colour, size, and even vascularity on the surface of the lesion [17]. Moreover, given increased specificity eNPM-DM versus experts, eNPM-DM may also help achieve better heath economics in NPC screening [33], simultaneously improving diagnostic accuracy and screening productivity. Furthermore, the combination of eNPM-DM and experts further increased the accuracy rate and decreased the false positive rate of NPC, identifying as many cases of malignancies as possible with minimal health expenditure in NPC screening. Accordingly, the emerging deep learning could serve as a powerful assistant in clinical practice, increasing the accuracy of screening, reducing cognitive burden on clinicians, positively impacting patients’ outcome and quality of life by fostering early intervention and reducing screening costs.
Additionally, our study offers a comprehensive method that is explicitly designed to develop a tool to segment nasopharyngeal malignancies in endoscopic images based on deep learning, which could be a promising biopsy guidance tool for nasopharyngeal malignancies, with the aim of increasing NPV of biopsy for malignancies. Here, eNPM-DM exhibited encouraging results in recognizing malignant areas in nasopharyngeal endoscopic images, which is consistent with the malignant lesion outlined by the experts. Accordingly, eNPM-DM could serve as a powerful biopsy guidance tool for resident oncologists or community physicians regardless of their limited experience in nasopharyngeal diseases.
To publicize our experience in nasopharyngeal malignancy detection and make full use of the advanced tool in clinical practice, we established an on-line platform (http://nasoai.sysucc.org.cn/). Both the patients and physicians may use this platform to assess the probability of malignancy in a certain image by uploading eligible nasopharyngeal endoscopic images to the artificial intelligence platform. If the lesion is recognized as malignant, the suggestive region for biopsy is provided.
There are limitations in this study. Given that all images were acquired from a single tertiary care centre in a highly endemic area of NPC, the diversity of nasopharyngeal diseases presented in this context might be reduced, subsequently resulting in over-fitting. However, the training curve revealed that the loss of the training was similar to that of the validation, which is indicative of a well-fit curve. In addition, NPC was the most prominent malignancy in this study, which might reduce the detection efficiency for other malignancies. However, given that NPC is the most common malignancy in the nasopharynx [1, 2] and the sensitivity of eNPM-DM in detecting nasopharyngeal malignancies was 90.2%, we believe that eNPM-DM is the most powerful auxiliary diagnosis tool in nasopharyngeal malignancy detection to date. One possible improvement could be a further increase in the spectrum of nasopharyngeal malignancies through collaboration with other centres in the future. Additionally, physical examination findings and laboratory test results, such as plasma antibody titters of EBV; and magnetic resonance imaging features of the nasopharynx and neck [9] can be taken into account during diagnosis in clinical practice. Therefore, integration of the endoscopic images, laboratory examination and radiologic images should be considered in nasopharyngeal malignancy detection based on deep learning. Similar to other deep leaning models, the exact features of eNPM-DM in malignancy detection remain unknown, and further investigation of detailed mechanisms is warranted. Particularly, since the model was trained on images, eNPM-DM could only render a diagnosis based on endoscopic images obtained in advance rather than real-time operation or video, and there is also a long and arduous way to combine eNPM-DM and the endoscopy system. Here, we manually selected 28,966 qualified images from numerous images and discarded the remaining images that are of poor quality or irrelevant. In future work, we plan to improve the performance of the model in image detection, identify the irrelevant images and evaluate image quality automatically. Finally, we plan to extend the developed deep learning image analysis framework to endoscopic image analysis and assessment in other types of cancers, such as gastric cancer, cervical cancer, and throat carcinoma.