Deep learning-based multi-class classification of thyroid disorders on Tc-99m scintigraphy using modified DenseNet-201.

OBJECTIVE

The thyroid gland plays a vital role in human body functions, including metabolism, and any dysfunction of the organ may exacerbate health risks. Early recovery from these conditions depends on the accurate diagnosis of accurate type of thyroid gland disorder. This study aims to develop an automated system for assisting physicians in clinical diagnosis of thyroid gland disorders.

METHOD: The transfer learning capability of the deep neural network model DenseNet-201 was leveraged, and the model was tailored by altering its fully connected layer and the classification layer for the classification of thyroid gland conditions into seven categories, namely cold nodule, hot nodule, multi-nodular goiter, nodular goiter, thyroiditis, toxic diffuse goiter and normal.

RESULTS

The class-wise and overall performance of the method was evaluated by computing quantitative metrics like accuracy, specificity, precision, sensitivity, F1-score, and area under the curve (AUC). The degree of similarity between the true labels of thyroid disorders annotated by experts and those predicted by the method was gauged using the kappa coefficient. For the five-fold cross-validated experimental results, we obtained an accuracy of 91.48 ± 2.79%, specificity of 98.58 ± 0.47%, precision of 91.57 ± 2.76%, sensitivity of 91.48 ± 2.77%, F1-score of 91.38 ± 2.82%, and AUC of 0.988 ± 0.006. Additionally, the degree of similarity in the diagnostic capability of the proposed method and medical experts was measured by computing the kappa coefficient as 0.9148.

CONCLUSION

The experimental results of the proposed method were compared with contemporary methods and illustrate relatively better performance in terms of accuracy, sensitivity, precision, and F1-score. The value of the kappa coefficient, 0.9148, also depicts that the proposed method has the potential for applicability in clinical diagnosis to assist physicians in assessing the accurate type of thyroid disorders.