Computer-aided mammography classification of malignant mass regions and normal regions based on novel texton features

Xizhao Li, Simon Williams, Gobert Lee, Min Deng

    Research output: Contribution to conferencePaperpeer-review

    4 Citations (Scopus)

    Abstract

    In this paper, new computer aided techniques for classifying malignant and normal mass regions with the novel texton features were proposed. Generally, the whole classifier consists of two stages: training and testing. In the first stage, all training images from the whole breasts were filtered by one of the maximum response filter banks (38 filters together but choose only the maximum 8 filter responses from each direction). Then mass regions were segmented automatically from those filtered images and aggregated together as the input of K-means clustering. This resulted in the generation of the final texton dictionary. Then, each region of interest (ROI) was modeled into the texton distributions. Similarly, in the second stage, all the novel testing images are represented by texton histograms of their own ROIs. In the end, Fisher Classifier was used for the classification and Receive Operation Characteristic (ROC) Curve was applied to show the performance (in the form of Az scores). Additionally, intensity independent texture analysis proposed in our previous research was used to normalize original images before the filter bank application. Performance turned out to be promising for the future research.

    Original languageEnglish
    Pages1431-1436
    Number of pages6
    DOIs
    Publication statusPublished - 1 Dec 2012
    EventICARV 2012 -
    Duration: 5 Jul 2012 → …

    Conference

    ConferenceICARV 2012
    Period5/07/12 → …

    Keywords

    • filter bank
    • Fisher Classifier
    • malignant
    • mammography classification
    • normal
    • ROC curve
    • texton

    Fingerprint

    Dive into the research topics of 'Computer-aided mammography classification of malignant mass regions and normal regions based on novel texton features'. Together they form a unique fingerprint.

    Cite this