Figure 1.

Pathways of experiments. Our experimental pathways mainly included three parts. (A) Data desensitization, removal of the sections of the patient’s personal information in the images. (B) Training and validation of ensemble learning for classification of thyroid nodules. In the segmentation part, the nodule area was manually marked and used to train the segmentation model. ROI and mask were extracted by the segmentation model. Then, three weak models were trained and combined to obtain an advanced classification model. (C) Comparison experiments with radiologists and other deep learning models, and external validation experiment. We then compared performance of the classification model with that of three ultrasound radiologists and four state-of-the-art deep learning models. Finally, we conducted an external validation using an independent dataset.