Diagnosis of thyroid disease using deep convolutional neural network models applied to thyroid scintigraphy images: a...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Nuclear Medicine Diagnostics
• Thyroid Disease Classification

Background:

• Accurate diagnosis of thyroid diseases is crucial for effective patient management.
• Nuclear medicine physicians interpret single-photon emission computed tomography (SPECT) images for thyroid evaluation.
• Improving the diagnostic performance in thyroid disease classification remains an area of active research.

Purpose of the Study:

• To enhance the diagnostic capabilities of nuclear medicine physicians in identifying thyroid diseases.
• To develop and validate a deep convolutional neural network (DCNN) model using SPECT imaging data.
• To compare the diagnostic performance of the DCNN model against human experts.

Main Methods:

• A multicenter retrospective study involving 3194 SPECT thyroid images for training and validation.
• Four pretrained DCNN models (AlexNet, ShuffleNetV2, MobileNetV3, ResNet-34) were evaluated for classifying thyroid diseases.
• The best-performing model (ResNet-34) underwent cross-validation and comparison with junior and senior nuclear medicine physicians.

Main Results:

• All four DCNN models achieved an accuracy greater than 0.85 in classifying SPECT thyroid images.
• The ResNet-34 model demonstrated superior performance with an accuracy of 0.944.
• The ResNet-34 model significantly outperformed senior physicians in both internal (94.4% vs. 85.6%) and external (93.1% vs. 86.8%) validation datasets.

Conclusions:

• Deep convolutional neural network models show high performance in diagnosing thyroid diseases from SPECT images.
• The DCNN model exhibited superior sensitivity and specificity compared to nuclear medicine physicians for various thyroid conditions.
• Deep learning models hold significant potential to augment clinicians' diagnostic efficiency in nuclear medicine.