Compression of Multimodal Biomedical Images using Neural Networks

Funder

Organizational Unit

Publications

LOSSY COMPRESSION OF BIOMEDICAL IMAGES FOR COMPUTER VISION ANALYSIS

Publication . Paulo, Edgar da Silva; Faria, Sérgio Manuel Maciel de; Távora, Luís Miguel de Oliveira Pegado de Noronha e; Thomaz, Lucas Arrabal

The exponential increase in medical and biomedical data acquisition is compelled by technological advances, namely in the imaging field. However, this exponential growth brings with it challenges in terms of processing capacity, transmission, and data storage. In response to this growing demand, increasingly efficient solutions have emerged, especially through computer vision for automatic image analysis and compression algorithms. This dissertation aims, on the one hand, to evaluate the performance of computer vision systems on previously compressed biomedical images. On the other hand, it increases the useful range of image variations, almost lossless and lossy, decreasing the impact of the change added by this method on the performance of computer vision algorithms in biomedical image analysis. In this sense, YOLO and Detectron2 are employed to evaluate the impact of coding distortion on their ability to detect mitochondria in electron microscopy images. The results of this study reveal that although the distortion introduced by compression affects their detection performance, it is negligible at lower compression ratios. Furthermore, two proposals are presented to improve the useful compression ratio, keeping the images characteristics that allow to perform the automatic detection of mitochondria. On the one hand, it is demonstrated that the proposed training methodology, which incorporates compressed versions of the original data during training, mitigates the impact of distortion on the performance of computer vision algorithms; on the other hand, allocating higher quality levels to regions of interest, compared to background elements, helps to sustain high performance at compression rates where computer vision algorithms typically start to lose effectiveness. These approaches allow the extension of the compression range with little impact on detection performance, thus contributing to the improvement of data processing, storage, and transmission in biomedical applications.

2024-06-14Master thesis

Open access

PROCESSING OF MICROSCOPY IMAGES IN THE COMPRESSED DOMAIN

Publication . Vasconcellos, Nicolas David Freire; Thomaz, Lucas Arrabal; Távora, Luís Miguel de Oliveira Pegado de Noronha e; Grilo, Carlos Fernando de Almeida; Miragaia, Rolando Lúcio Germano

This work presents a novel approach to the detection of mitochondria in microscopy images within the compressed domain, eliminating the need for full image decompression while maintaining high detection accuracy. The proposed architecture, Processing of Microscopy Images in the Compressed Domain (ProMIC), integrates a Domain Translator architecture, enabling object detection networks to operate directly on images’ latent representations. This approach significantly reduces computational costs while preserving essential visual information for detection tasks. The study also included a preliminary analysis of JPEG-AI compression efficiency against classical codecs (HEVC/H.265 and VVC/H.266) across multiple datasets. The proposed method builds upon a Base model inspired by a previous compresseddomain object detection approach. Key innovations in this work include: the use of a Lightweight Residual Block that enhances feature extraction from latent representations, improving detection robustness across different compression rates; a Pixel Domain Fine-Tuning process that enhances the Domain Adaptation of the model; and a Guided Domain Translator Training strategy that refines the Domain Translator to extract meaningful information directly from latent codes, minimising the gap between compressed and pixel-domain object detection models. Extensive experiments were conducted to compare the performance of both the Base model and the proposed architecture against a reference object detection model that operates on reconstructed images. Results on the Lucchi++ dataset show that ProMIC significantly outperforms the Base model. Notably, the proposed architecture exceeds the reference model in mean average precision, achieving a 2.31 percentage point improvement at the highest bitrate. This highlights the potential of compressed-domain image processing to capture meaningful features that may be lost in fully reconstructed images. Additionally, applying the ProMIC method reduced the complexity of the object detection system by 42.34% compared to the reference model. These findings validate the feasibility of direct object detection in latent codes, offering a computationally efficient alternative to traditional pixel-domain approaches. The proposed framework has promising applications in real-time processing, embedded systems, and large-scale AI-driven image analysis, paving the way for more efficient and scalable vision models that operate directly in the compressed domain.

2025-03-19Master thesis

Open access