2.2 Data Preparation
The dataset was randomly split into two parts: 1450 images for training and 729 images for testing. For training, the number of images in the malignant and benign groups was 199 and 1251, respectively. For the testing group, 89 were malign and 640 were benign (Table 1). Two senior otolaryngologists used the labeling software Label Tool to interpret the laryngoscope images and label the exact region of the biopsy.