Loading...
Publication
ROI -BASED CODING OF BIOMEDICAL IMAGES FOR MACHINE ANALYSIS
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 12.74 MB | Adobe PDF |
Authors
Abstract(s)
The increasing volume of data acquired and generated daily in the healthcare sector,
driven by technological advancements, brings significant benefits to patient diagnosis
and research. However, this growth also presents considerable challenges in the
analysis and processing of such data. To address these difficulties, computer vision
algorithms have emerged as powerful tools, capable of automating repetitive and
time-consuming tasks, enabling faster and more accurate processing.
At the same time, the growing volume of data places pressure on storage and
transmission capabilities, demanding efficient compression methods to minimise
its size. In the literature, various approaches are found, primarily divided into
two categories: lossy and lossless compression. While lossless methods ensure data
integrity, they do not achieve compression rates as high as lossy algorithms. The
latter, despite significantly reducing file sizes, introduces distortions that may
compromise image quality, affecting the accuracy of automated systems.
This dissertation focuses on two main challenges: first, evaluating the impact of
image compression on the performance of biomedical computer vision systems, and
second, improving compression efficiency without compromising the accuracy of
these algorithms. To this end, detection and segmentation models, such as YOLOv8
and SAM, were used to analyse the effect of distortion caused by encoding on the
localisation and segmentation of mitochondria in two datasets of electron microscopy
images.
To enhance model performance at higher compression levels, two methodologies
were implemented. The first focuses on domain adaptation, fine-tuning the models
to recognise and compensate for distortions introduced by compression, specifically
in HEVC/H.265 and VVC/H.266 encoders. The second approach proposes contentaware
encoder adaptation, allowing the assignment of different quality levels to
selected regions of interest. This method aims to reduce storage and bandwidth
requirements without significantly compromising the performance of deep learningbased
models.
Experimental results demonstrate that region-of-interest-based encoding strategies
effectively reduce compression rates while maintaining model accuracy. In particular,
the proposed methodologies allowed to achieve an average performance improvement
of up to 23.70% for the same bpp range and a data size reduction of up to 74.96%.
Additionally, a Pareto-based optimisation algorithm was proposed to determine the most suitable encoding configurations for different standards and models, ensuring
a balance between compression efficiency and object detection performance.
Description
I acknowledge the financial support provided by the Fundação para a Ciência e a Tecnologia (FCT), Portugal under projects CoMBINNe 2022.09914.PTDC (DOI:10.54499/2022.09914.PTDC), Programa Operacional Regional do Centro, and by FCT/MCTES through national funds and when applicable co-funded by EU funds under the project UIDB/EEA/50008/2020 (DOI: 10.54499/UIDB/50008/2020) and
LA/P/0109/2020 (DOI: 10.54499/LA/P/0109/2020).
Keywords
Imagiologia biomédica Microscopia eletrónica Compressão com Perdas YOLO SAML HEVC VVC Codificação de Regiões de Interesse Preservação de Conteúdo