Evaluation of Dimensionality Reduction Techniques for Thyroid Cancer Diagnosis Using Ultrasound and Histopathological Images

The thyroid is a vital organ that helps the appropriate functioning of the human body. The cases of thyroid cancer are increasing daily, and early diagnosis may help medical practitioners reduce the healthcare burden. The recent advancement in AI-based computed-aided diagnosis tools has shown promising results in different healthcare domains but still requires improvement in their performances. In this study, we evaluate the performance of different dimensionality reduction techniques to improve the performance of the thyroid diagnosis model. A pre-trained Swin transformer model is utilized for feature extraction from ultrasound and histopathological thyroid images. Six dimensionality reduction techniques (Principal Component Analysis, Truncated Singular Value Decomposition, Fast Independent Component Analysis, Isometric Feature Mapping, Locally Linear Embedding and Uniform Manifold Approximation and Projection techniques) are employed for feature reduction in wrapper-based feature selection mode. The logistic regression classifier is used to assess the performance of the proposed dimensionality reduction techniques in wrapper feature reduction mode. The dimensionality reduction techniques are ranked using the MEREC-TOPSIS method, where performance metrics are evaluated on the test datasets. The locally linear embedding technique ranked as the best dimensionality reduction technique for the features obtained from both datasets. The results for the ultrasound dataset include an accuracy of 0.9161, an AUC-ROC value of 0.9612, and an F2-Score of 0.9028. In contrast, when considering the histopathological dataset, the highest values gained for accuracy, AUC-ROC, and F2-score are 0.8223, 0.8523, and 0.8357, respectively. The model based on the proposed dimensionality reduction technique achieved better performances with less computational complexities due to reduced feature space.