Abstract
Content-based medical image retrieval (CBMIR) is an approach utilized for extracting pertinent medical images from extensive databases by focusing on their visual attributes instead of relying on textual information. This method entails examining the visual qualities of medical images, including texture, shape, intensity, and spatial relationships, in order to detect resemblances and patterns. In this study, the analysis focuses on six prominent low-level handcrafted feature techniques and eight transfer learning with pre-trained deep learning models for extracting features for CBMIR systems. Image indexing is crucial in CBMIR systems, particularly with high-dimensional data and the extremely sparse distribution of original data called the 'curse of dimensionality' problem. To address such problem, we use Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE) for dimensionality reduction. This experiments are performed on two benchmark datasets: Medical MNIST and KVASIR. For Medical MNIST datasets, handcrafted features are effective for distinct texture characteristics that are easily discernible to the human eye, however deep learning approaches are necessary for datasets with smaller shapes, sizes, and textures, like KVASIR dataset to minimize the semantic gap. The performance of the feature based techniques is evaluated using metrics: Precision, Recall, and F1-score. The handcrafted technique with t-SNE maintains constant performance with maximum 99.89% fewer dimensions compared to the full-featured technique. And with KVASIR dataset, using DCNN architecture with t-SNE, we achieve a maximum dimensionality reduction of 75% while maintaining consistent results.