A prediction of late cervical lymph node metastasis by ultrasound images of tongue cancer using deep learning method

Abstract

ObjectiveOccult cervical lymph node metastasis is an important prognostic factor in squamous cell carcinoma of the tongue and is related to factors such as depth of invasion and angiogenesis. The objective of this study was to evaluate the performance of deep learning for predicting occult cervical metastasis from brightness mode (B-mode) ultrasound images of early-stage tongue cancer. Methods50 patients with early-stage tongue cancer who underwent ultrasound guided partial glossectomy at the Department of Oral Surgery at Saitama Cancer Center between 2015 and 2019 were divided into two groups: those with metastasis (n = 22) and those without metastasis (n = 28). B-mode ultrasound images of each patient's tongue cancer were cropped to a 128 × 128-pixel square, and images were assigned to training or testing data in an 8:2 ratio. The training iteration generation (number of epochs) was set to 120 using Neural Network Console software (Sony Corporation), and the resulting learning model was produced. Testing data were input into the learning model to evaluate its performance in predicting metastasis. ResultsPrediction of occult cervical lymph node metastasis by deep learning using B-mode ultrasound images had 79.8 % accuracy, 78.4 % sensitivity, 80.8 % specificity, 74.4 % positive predictive value, and 84.0 % negative predictive value. ConclusionThese findings suggest that deep learning using ultrasound images is useful for predicting postoperative cervical lymph node metastasis from tongue cancer.