Retrieval System in Web-based Platform with Limited Resource Environment Using Compact Features of Medical Images

A Web Content-Based Image Retrieval (CBIR) System can be utilized as a training platform for users such as medical students and resident doctors, helping them gain experience in diagnosing through medical images. Based on misdiagnosed image by the users, the CBIR system will find similar images to be provided as training feedback, allowing them to understand diagnostic errors and refine their skills. Deploying such CBIR systems on web platforms is advantageous for ease of accessibility but imposes some limitation due to the lack of computation power. Utilizing handcrafted algorithms to extract features are suitable due to their low computation requirements, but they often compromise in retrieval performance in medical images. Convolutional Neural Networks (CNN), on the other hand, have been proven to yield better performance in feature extraction and image retrieval tasks on medical images. However, CNN-based approaches require significant computational resources and Graphical Processing Unit (GPU), making them less ideal for deployment in web servers that often lack GPU, and servers equipped with GPU are costly. Furthermore, CNN-based techniques produce high number of features, requiring significant storage and computation, often needing further feature selection process, adding more complexity to the computations and affecting responsiveness. This study aims to propose Compact CBIR (ComCBIR), a compact and efficient CBIR system designed for web-based platforms with limited resources. ComCBIR simultaneously extracts and reduces feature size within a single forward pass of a CNN. This is achieved by modifying CNN architecture and fine-tuning using pairwise Contrastive Loss to minimize the feature representation distance between similar image while maximizing dissimilar image distance. The Representational Learning approach result in feature that are learned directly from the training process and allow for more compact feature size compared to classifier-based, which typically requires a larger latent feature size. Combined with mobile focused network, this approach results in a system that are suitable in web-based platform, achieving faster image retrieval response and efficient storage utilization. Trained on 14.000 endoscopic and dermatoscopic images, experiments demonstrate notable image retrieval results, with mean average precision (mAP) scores of 0.923 for endoscopic images and 0.809 for dermatoscopic images. These retrieval results are obtained using feature size of only eight, with performance comparable to using higher feature count. Furthermore, the system efficiently stores features for 14,000 images in just 1.75 MB. This demonstrates the potential of ComCBIR as an effective solution for medical image retrieval in resource-constrained environments.