Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • 33 views

    Poster - Israel Chaparro Cruz

    Uploaded by

    Brando Efrain Marca Machaca
    The document proposes a novel GAN generator called BORDE for multimodal semantic image synthesis of brain images. BORDE introduces a new normalization block that injects a-priori information like boundaries and sub-regions directly into the normalization layer. This helps generate high-quality and semantically coherent synthetic brain images. Experimental results show BORDE outperforms SOTA baselines like SPADE and SEAN on brain tumor and stroke segmentation tasks, demonstrating the need for domain-specific techniques in medical image GANs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Poster - Israel Chaparro Cruz

    Uploaded by

    Brando Efrain Marca Machaca
    33 views1 page
    The document proposes a novel GAN generator called BORDE for multimodal semantic image synthesis of brain images. BORDE introduces a new normalization block that injects a-priori information like boundaries and sub-regions directly into the normalization layer. This helps generate high-quality and semantically coherent synthetic brain images. Experimental results show BORDE outperforms SOTA baselines like SPADE and SEAN on brain tumor and stroke segmentation tasks, demonstrating the need for domain-specific techniques in medical image GANs.

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document proposes a novel GAN generator called BORDE for multimodal semantic image synthesis of brain images. BORDE introduces a new normalization block that injects a-priori information like boundaries and sub-regions directly into the normalization layer. This helps generate high-quality and semantically coherent synthetic brain images. Experimental results show BORDE outperforms SOTA baselines like SPADE and SEAN on brain tumor and stroke segmentation tasks, demonstrating the need for domain-specific techniques in medical image GANs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    33 views1 page

    Poster - Israel Chaparro Cruz

    Uploaded by

    Brando Efrain Marca Machaca
    The document proposes a novel GAN generator called BORDE for multimodal semantic image synthesis of brain images. BORDE introduces a new normalization block that injects a-priori information like boundaries and sub-regions directly into the normalization layer. This helps generate high-quality and semantically coherent synthetic brain images. Experimental results show BORDE outperforms SOTA baselines like SPADE and SEAN on brain tumor and stroke segmentation tasks, demonstrating the need for domain-specific techniques in medical image GANs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 1

    Israel N. Chaparro-Cruza and Javier A.

    Montoya-Zegarraa,b

    a Department of Computer Science, Universidad Catolica San Pablo, Arequipa, Perú. b Institute for Biomedical Engineering, ETH Zurich, Zurich, Switzerland.

    Abstract: Medical images are often expensive to acquire. Thus, the use of medical datasets can be restrictive for training deep learning models. In this paper, we propose a novel GAN generator for multimodal semantic image synthesis of brain images based on a novel denormalization
    block named BORDE. Our model generates very realistic and plausible synthetic images that when combined with real data helps to improve the accuracy in brain segmentation tasks. This allows us to prove the need for more domain-specific techniques in GANs models.

    Introduction Related Work Proposal Experimental Results:


    Brain stroke represents the second leading cause of death One seminal work is Pix2pix [15], which comprises an BORDE Block: Given an a-priori input, we propose a block called BORDE to Quantitative results:
    worldwide. Gliomas are the most common tumor with various encoder-decoder generator that takes as input a generate realistic-looking brain images with very fine details characterized by
    heterogeneous histological sub-regions characterized by varying semantic map. Thenceforth, different architectures and precise contours and localized texture information.
    intensity profiles. Accurate segmentation is decisive for loss modifications have been proposed [10], [16], [17]
    diagnosis and treatment. to improve Pix2pix’s image synthesis quality.

    Deep learning advances have shown promising results but In SPADE [10], the authors noticed that the
    require a massive amount of annotated datasets; nonetheless, normalization layers tend to “wash away” the input
    especially in the medical domain, such annotations are expensive semantic information. To address this issue, the authors
    to acquire, require expert annotation level, and can be limited propose using the input semantic map to denormalize
    due to privacy issues. or modulate the model activations through a
    normalization layers.
    Motivation: GANs can be used to generate additional annotated
    training data. If cGANs were trained to generate realistic-looking Normalization layers helps to stabilize and speed up the
    images conditioned on semantic map, then the model can be learning process in two steps. A more recent approach
    used for semantic image synthesis tasks. This task is still is SEAN [11]. This model introduces a simple but
    challenging, many state-of-the-art models use scene, or face effective block conditioned on a segmentation mask In a BORDE block, the input information is split over different normalization
    benchmark datasets. that describes the semantic regions in the output image. layers. This helps to generate semantic-coherent synthetic images.
    We noticed that those datasets differ quite a lot from medical
    images datasets. BORDE Generator: Our generator relies on a decoder architecture built by
    Proposal BORDE blocks.

    Denormalization: We propose a novel denormalization


    technique called BOundary and sub-Region Qualitative Results:
    DEnormalization (BORDE) intended for brain image
    synthesis.

    Given the a-priori information M (boundary mask,


    foreground or sub-regions mask/map), the goal is to
    inject the a-priori information directly on the Noise is injected into the model in two different ways: (i) a constant Gaussian
    denormalization layer. input noise to represent the general sketch, (ii) different per channel noise
    auxiliary input. In addition, residual connections are added at each stage [25].
    Finally, our generator is multi-task generating images of the required modalities
    and also an image of the segmentation contours (boundary) where the foreground
    contour is also included.

    Methodology:
    Datasets: 1) BraTS 2020: The Multimodal Brain Tumor Segmentation Challenge
    Conclusions:
    We have proposed BORDE, a normalization block
    The denormalization process is given by the calculation (BraTS) contains multimodal MRI scans from 369 different patients [1], [14],
    and generator, that produce realistic multimodal
    of the parameters γ and β and applied to the normalized [28]. 2) ISLES 2018: The Ischemic Stroke Lesion Segmentation challenge
    brain images of tumor lesions or ischemic strokes.
    activation map considering two important (ISLES) contains 63 different multimodal CT perfusion scans [2], [29].
    The paper is available at (scan We further demonstrate that our model
    characteristics: (i) spatial-variance and (ii) instance- Pre-processing: We first convert the 3D volumes and their annotations into stacks
    QR code) and has been published outperforms different SOTA baselines.
    specific normalization. of 2D images and add the foreground as part of the segmentation map.
    at the 34th Conference on
    Baselines: : SPADE [10] and SEAN [11]
    Graphics, Patterns and Images - Acknowledgments:
    These properties allow improving the image quality, Evaluation: We train our proposal and baselines with both datasets to generate
    SIBGRAPI 2021. This study was supported by grant 234-2015-
    especially in smaller sub-regions, by incrementally synthetic images and use them as Data Augmentation (DA) in a segmentation
    Contact: ichaparroc@gmail.com injecting a-priori information network U-Net [30]. FONDECYT (Master Program) from
    CONCYTEC-PERU.

    You might also like