Benchmarking the use of Dimensional Reduction Techniques: A Case Study of Oesophageal Cancer Image Analysis

Abstract

The dimensionality reduction method is one of the most popular approaches for handling complex data characterised by numerous features and variables. In this work, we benchmarked the application of different techniques to interpret cancer-based in vivo microscopic images. We focus on several dimensionality reduction methods, including PCA, LDA, t-SNE, and UMAP, to evaluate the performance of the image dataset analysis (5043 images). The benchmarking study establishes the efficacy of traditional machine learning algorithms for biomedical image analysis. Model descriptions based on logistic regression, support vector, K-means clustering, K-nearest neighbour, random forest, gradient boosting, and adaboost classifiers were employed. The study also evaluates the importance of visualisation techniques relevant for identifying hidden patterns, anomalies, and trends that are not readily discernible in high-dimensional data. The benchmarking study uses approaches like random splits and K-fold cross-validation. Further evaluation metrics such as accuracy, sensitivity, specificity, and ROC-AUC score are employed to assess the performance of the employed dimensionality reduction methods. Their relevance for data visualisation as well as predictive modelling is demonstrated. Overall, the study is useful for understanding the relevance of effective data classification and visualisation problems, thereby enhancing the interpretability and analysis of biomedical images.