A preliminary application of intraoral Doppler ultrasound images to deep learning techniques for predicting late cervical lymph node metastasis in early tongue cancers. Issue 2 (26th November 2019)

• Record Type: Journal Article
• Authors: Ariji, Yoshiko; Fukuda, Motoki; Kise, Yoshitaka; Nozawa, Michihito; Nagao, Toru; Nakayama, Atsushi; Sugita, Yoshihiko; Katumata, Akitoshi; Ariji, Eiichiro
• Is Part Of: Oral science international. Volume 17:Issue 2(2020)
• Page Start: 59
• Page End: 66
• Publication Date: 2019-11-26
• Subjects: deep learning -- depth of invasion -- intraoral Doppler ultrasound images -- late cervical lymph node metastasis -- tongue cancer; Oral medicine -- Periodicals; Mouth -- Diseases -- Periodicals; Mouth -- Diseases; Oral medicine; Oral Medicine -- Periodicals; Stomatognathic Diseases -- Periodicals; Periodicals; 617.522
• DOI: 10.1002/osi2.1039
• Languages: English
• ISSNs: 1348-8643

Abstract

Abstract: Aims: Various factors, including depth of invasion (DOI) and hemodynamics have been linked with the prediction of late cervical lymph nodes metastasis in patients with early tongue cancers. The objective of this study was to examine the deep learning performance of the intraoral Doppler ultrasound images for predicting the late cervical metastasis, by comparing DOI. Methods: Thirty‐three patients with early squamous cell tongue carcinomas were divided into two groups: 12 with late cervical metastasis, and 21 without metastasis. Intraoral Doppler ultrasound images of all subjects were cropped to 400 × 400 pixel squares, and 80% were used for a training dataset, and 20% were used for a testing dataset. The training dataset was imported into the DIGITS deep learning training system, the learning process for 300 epochs was performed using AlexNet neural network, and the resultant learning model was created. The testing dataset was applied to the model to evaluate the performance for distinguishing between the two groups. Results: Use of intraoral Doppler ultrasound images for predicting the late cervical metastasis achieved deep learning performances of 0.883 for the area under the ROC curve (AUC), 85.9% for accuracy, and 84.0% for sensitivity. On the other hand, the corresponding performances of DOI were 0.873, 84.8%, and 75.0%, using a DOI threshold of 5.6 mm. Conclusion: Our findings suggested that the performance of a deep learning system using intraoral Doppler ultrasound images of early tongue cancers to predict late cervical metastasis was sufficiently high, suggesting possible applications in imaging diagnosis support.