Fusing deep and handcrafted features for intelligent recognition of uptake patterns on thyroid scintigraphy

Thyroid scintigraphy is an important investigation for the clinical diagnosis of thyroid diseases. Thyroid diseases often present characteristic abnormal patterns in scintigraphic images. In this study, we developed an automated recognition model for thyroid uptake patterns. These patterns were classified into six categories, and they are diffusely increased, diffusely decreased, heterogeneous, focally increased, focally decreased, and normal. This study is the first report on such data for automated thyroid pattern interpretation to the best of our knowledge. A thyroid uptake pattern recognition network (TPRNet) was developed, using deep and handcrafted features of thyroid patterns to perform classification. The method can be trained using the traditional back-propagation algorithm. An in-house dataset that contains 4263 thyroid scintigraphic images was constructed to train and validate the TPRNet. Furthermore, an external test dataset that includes 1318 images was constructed to test the TPRNet. Experimental results show that the proposed method is effective for recognizing thyroid patterns in scintigraphic images. It outperforms compared methods, notably in the external test dataset, demonstrating the good generalization ability of the TPRNet. The proposed method is also compared against four physicians' judgments on recognizing thyroid patterns, resulting in a performance that is comparable to that of experienced doctors, showing that it could be used in clinical practice. (c) 2021 Published by Elsevier B.V.

Automated summary

In this study, an automated recognition model for thyroid uptake patterns was developed and tested. The results showed that the proposed method is effective for recognizing thyroid patterns in scintigraphic images, surpassing other methods and comparable to experienced physicians' judgments.