Multiclass Classification of Abnormal Endocrine Gland States Using CNNs and Transformers

The thyroid gland plays a crucial role in maintaining the proper metabolic activities of the body. Any deviations in the production of thyroid hormone can result in impaired metabolism. Hence, early detection of the infected gland is critical to providing timely and appropriate treatment. Delay in treatment can worsen the condition of the patient. It is observed that women are more likely to be diagnosed with thyroid disorders than men. Over the past few decades, researchers have explored various techniques to automate the diagnosis of thyroid gland abnormalities. In this study, multilayer perceptron, support vector machine, and random forest are used to develop a computer-aided diagnosis system to detect thyroid gland abnormalities. Numerical input data are used with multilayer perceptron to perform multiclass classification of the classes such as hyperthyroid, hypothyroid, sick, and negative. The model showed promising results with a classification accuracy of 82% in the presence of noisy data, using 5-fold cross-validation. Support vector machine performed slightly less, with a classification accuracy of 80%. Both models exhibited good performance, with 85% precision and 95% recall for SVM and 88% precision and 91% recall for multilayer perceptron. This was the case when numerical input data is used. The accuracy of CNN model reached 99.9% with a loss of 4% on the validation data.