Learning domain-invariant representations of histological images

Onderzoeksoutput: Bijdrage aan tijdschriftTijdschriftartikelAcademicpeer review

Uittreksel

Histological images present high appearance variability due to inconsistent latent parameters related to the preparation and scanning procedure of histological slides, as well as the inherent biological variability of tissues. Machine-learning models are trained with images from a limited set of domains, and are expected to generalize to images from unseen domains. Methodological design choices have to be made in order to yield domain invariance and proper generalization. In digital pathology, standard approaches focus either on ad-hoc normalization of the latent parameters based on prior knowledge, such as staining normalization, or aim at anticipating new variations of these parameters via data augmentation. Since every histological image originates from a unique data distribution, we propose to consider every histological slide of the training data as a domain and investigated the alternative approach of domain-adversarial training to learn features that are invariant to this available domain information. We carried out a comparative analysis with staining normalization and data augmentation on two different tasks: generalization to images acquired in unseen pathology labs for mitosis detection and generalization to unseen organs for nuclei segmentation. We report that the utility of each method depends on the type of task and type of data variability present at training and test time. The proposed framework for domain-adversarial training is able to improve generalization performances on top of conventional methods.
TaalEngels
Artikelnummer162
Aantal pagina's11
TijdschriftFrontiers in Medicine
Volume6
DOI's
StatusGepubliceerd - 16 jul 2019

Trefwoorden

    Citeer dit

    @article{51de4585722247878597edcf86f47da1,
    title = "Learning domain-invariant representations of histological images",
    abstract = "Histological images present high appearance variability due to inconsistent latent parameters related to the preparation and scanning procedure of histological slides, as well as the inherent biological variability of tissues. Machine-learning models are trained with images from a limited set of domains, and are expected to generalize to images from unseen domains. Methodological design choices have to be made in order to yield domain invariance and proper generalization. In digital pathology, standard approaches focus either on ad-hoc normalization of the latent parameters based on prior knowledge, such as staining normalization, or aim at anticipating new variations of these parameters via data augmentation. Since every histological image originates from a unique data distribution, we propose to consider every histological slide of the training data as a domain and investigated the alternative approach of domain-adversarial training to learn features that are invariant to this available domain information. We carried out a comparative analysis with staining normalization and data augmentation on two different tasks: generalization to images acquired in unseen pathology labs for mitosis detection and generalization to unseen organs for nuclei segmentation. We report that the utility of each method depends on the type of task and type of data variability present at training and test time. The proposed framework for domain-adversarial training is able to improve generalization performances on top of conventional methods.",
    keywords = "histopathology image analysis, domain-invariant representation, domain-adversarial convolutional network, mitosis detection, nuclei segmentation",
    author = "Lafarge, {Maxime W.} and Pluim, {Josien P. W.} and Eppenhof, {Koen A. J.} and Mitko Veta",
    year = "2019",
    month = "7",
    day = "16",
    doi = "10.3389/fmed.2019.00162",
    language = "English",
    volume = "6",
    journal = "Frontiers in Medicine",
    issn = "2296-858X",
    publisher = "Frontiers Media",

    }

    Learning domain-invariant representations of histological images. / Lafarge, Maxime W. (Corresponding author); Pluim, Josien P. W.; Eppenhof, Koen A. J.; Veta, Mitko.

    In: Frontiers in Medicine, Vol. 6, 162, 16.07.2019.

    Onderzoeksoutput: Bijdrage aan tijdschriftTijdschriftartikelAcademicpeer review

    TY - JOUR

    T1 - Learning domain-invariant representations of histological images

    AU - Lafarge,Maxime W.

    AU - Pluim,Josien P. W.

    AU - Eppenhof,Koen A. J.

    AU - Veta,Mitko

    PY - 2019/7/16

    Y1 - 2019/7/16

    N2 - Histological images present high appearance variability due to inconsistent latent parameters related to the preparation and scanning procedure of histological slides, as well as the inherent biological variability of tissues. Machine-learning models are trained with images from a limited set of domains, and are expected to generalize to images from unseen domains. Methodological design choices have to be made in order to yield domain invariance and proper generalization. In digital pathology, standard approaches focus either on ad-hoc normalization of the latent parameters based on prior knowledge, such as staining normalization, or aim at anticipating new variations of these parameters via data augmentation. Since every histological image originates from a unique data distribution, we propose to consider every histological slide of the training data as a domain and investigated the alternative approach of domain-adversarial training to learn features that are invariant to this available domain information. We carried out a comparative analysis with staining normalization and data augmentation on two different tasks: generalization to images acquired in unseen pathology labs for mitosis detection and generalization to unseen organs for nuclei segmentation. We report that the utility of each method depends on the type of task and type of data variability present at training and test time. The proposed framework for domain-adversarial training is able to improve generalization performances on top of conventional methods.

    AB - Histological images present high appearance variability due to inconsistent latent parameters related to the preparation and scanning procedure of histological slides, as well as the inherent biological variability of tissues. Machine-learning models are trained with images from a limited set of domains, and are expected to generalize to images from unseen domains. Methodological design choices have to be made in order to yield domain invariance and proper generalization. In digital pathology, standard approaches focus either on ad-hoc normalization of the latent parameters based on prior knowledge, such as staining normalization, or aim at anticipating new variations of these parameters via data augmentation. Since every histological image originates from a unique data distribution, we propose to consider every histological slide of the training data as a domain and investigated the alternative approach of domain-adversarial training to learn features that are invariant to this available domain information. We carried out a comparative analysis with staining normalization and data augmentation on two different tasks: generalization to images acquired in unseen pathology labs for mitosis detection and generalization to unseen organs for nuclei segmentation. We report that the utility of each method depends on the type of task and type of data variability present at training and test time. The proposed framework for domain-adversarial training is able to improve generalization performances on top of conventional methods.

    KW - histopathology image analysis

    KW - domain-invariant representation

    KW - domain-adversarial convolutional network

    KW - mitosis detection

    KW - nuclei segmentation

    U2 - 10.3389/fmed.2019.00162

    DO - 10.3389/fmed.2019.00162

    M3 - Article

    VL - 6

    JO - Frontiers in Medicine

    T2 - Frontiers in Medicine

    JF - Frontiers in Medicine

    SN - 2296-858X

    M1 - 162

    ER -