Optimized Deep Features for Colon Histology Image Retrieval

Recently, deep learning techniques have been widely used in content-based image retrieval (CBIR) due to their success in bridging the "semantic gap" issue. Nevertheless, the high-dimensional features extracted through deep learning models usually lead to a high computational cost. This issue has a major impact especially when it involves the retrieval of medical images, where a short response time is very important. To address this issue, we propose in this paper an effective optimization approach for reducing the dimension of deep features extracted from colon histology images. Specifically, to extract the deep features, we first apply transfer learning on the pre-trained ResNet18 and GoogLeNet networks. Then, we use sine cosine algorithm (SCA) to optimize the extracted and combined features. Experiments conducted on Kather-5k colorectal cancer histology image dataset demonstrate the effectiveness of the proposed dimensionality reduction method in reducing the feature dimension, with a gain up to 50% better, while keeping good retrieval performance.