Deep Learning Based Multiclass Classification of Thyroid Scintigraphy

Introduction: Thyroid has many abnormalities, and its timely and correct diagnosis can help the patient's recovery process. Differentiation between normal cases and abnormal ones, including thyroiditis and goiter diseases, was performed by multi-modality including ultrasonography, computed tomography (CT), and thyroid scintigraphy (TS) with ^99mTc-pertechnetate. Since thyroid scintigraphy is a functional imaging method, it is efficient to diagnose goiter and thyroiditis. Convolutional neural network (CNN) model shows good performance in the classification of the most common disorders. In this study, we aim to use CNN deep learning algorithms in detecting the normal thyroid, thyroiditis, and goiter using a thyroid scan

Methods: In the current study, 239 patients from two nuclear medicine centers were classified into three categories: normal (55), thyroiditis (97), and goiter (87). TS was performed by one head gamma camera with 128×128 matrix size after 20-30 minute of intravenously injection of 370 MBqs ^99mTc-pertechnetate. Patients were split into two groups training (80%) and test (20%) datasets. Fifteen percent of the training dataset was used for validation. We used two CNN models, including simple CNN with three convolutional layer and ResNet152V2. Instead of the top layer for ResNet152V2, we used two dense layers (512 and 128), and batch normalization and ReLU activation function for each layer. Softmax layer at the end of the models was used for multiclass classification. Models were trained with 100 epochs with early stopping and 16 batch sizes. Accuracy, precision, recall, and f1-score were used for model performance metrics on test dataset.

Results: ResNet152V2 model had a best performance with accuracy of 0.77. Recall (0.73, 0.85 and .71), precision (0.67, 0.85 and 0.75), f1-score (0.70, 0.85 and 0.73), were achieved for normal, thyroiditis and goiter, respectively by using ResNet152V2. Simple CNN model had accuracy of 0.73. Recall (0.55, 0.95 and 0.59), precision (0.60, 0.86 and 0.62), f1-score (0.57, 0.90 and 0.61), were achieved normal, thyroiditis and goiter, respectively by using simple CNN.

Conclusions: Results of the current study show that the feasibility of diagnosis thyroid abnormality by using deep learning models. The proposed method can effectively classify different thyroid scans into normal, thyroiditis, and goiter groups. Three classes were obtained, which provide more insight into the diagnosis of the thyroid that can be used for diagnostic and prognostic purposes in clinical routine.

• Download figure
• Open in new tab
• Download powerpoint