Endocrine Abstracts

Evaluation of learning methods similar to deep learning and device using deep learning for the diagnosis of thyroid nodules.BackgroundWe recently developed a deep convolutional neural network algorithm (SEveRance Artificial intelligence program, SERA) using 13,560 ultrasonography images of thyroid nodules labeled benign and malignant and this algorithm showed comparable diagnostic performance with experienced radiologists. We wondered whether this self-learning method of SERA could be adapted for human learning as an ancillary approach to man-to-man training.

Methods: Twenty-one internal medicine residents studied the “learning set” in three replicates which was composed of 3,000 images of selected from 13,560 thyroid nodules and their diagnostic performances were evaluated before study and after every learning session using the “test set” which was composed of 120 thyroid nodule images. The diagnostic performances of eight radiology residents were evaluated before and after man-to-man training using the same “test set”. After final test, all readers once again evaluated the “test set” with the assistance of SERA.

Results: Before study, the mean area under the receiver operating characteristic (AUROC) of internal medicine residents were considerably lower than that of radiology residents (0.578 and 0.701, respectively). Diagnostic performance of internal medicine residents, although not as much as radiology residents who received man-to-man training (AUROC = 0.735), increased over the course of the learning program (AUROC = 0.665, 0.689, and 0.709, respectively). All diagnostic performances of internal medicine residents and radiology residents were better with the assistance of SERA (AUROC 0.755 and 0.768, respectively).

Conclusion: A novel iterative learning method using selected ultrasound images from big data sets can help beginners learn to differentiate between benign and malignant thyroid nodules. With the assistance of SERA, the diagnostic performances of readers with various experiences in thyroid imaging could be further improved.