39-Issue 3

Permanent URI for this collection

https://diglib.eg.org/handle/10.2312/2632882

Browse

Browsing 39-Issue 3 by Subject "Digital libraries and archives"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Knowledge-Assisted Comparative Assessment of Breast Cancer using Dynamic Contrast-Enhanced Magnetic Resonance Imaging
    (The Eurographics Association and John Wiley & Sons Ltd., 2020) Nie, Kai; Baltzer, Pascal; Preim, Bernhard; Mistelbauer, Gabriel; Viola, Ivan and Gleicher, Michael and Landesberger von Antburg, Tatiana
    Breast perfusion data are dynamic medical image data that depict perfusion characteristics of the investigated tissue. These data consist of a series of static datasets that are acquired at different time points and aggregated into time intensity curves (TICs) for each voxel. The characteristics of these TICs provide important information about a lesion's composition, but their analysis is time-consuming due to their large number. Subsequently, these TICs are used to classify a lesion as benign or malignant. This lesion scoring is commonly done manually by physicians and may therefore be subject to bias. We propose an approach that addresses both of these problems by combining an automated lesion classification with a visual confirmatory analysis, especially for uncertain cases. Firstly, we cluster the TICs of a lesion using ordering points to identify the clustering structure (OPTICS) and then visualize these clusters. Together with their relative size, they are added to a library. We then model fuzzy inference rules by using the lesion's TIC clusters as antecedents and its score as consequent. Using a fuzzy scoring system, we can suggest a score for a new lesion. Secondly, to allow physicians to confirm the suggestion in uncertain cases, we display the TIC clusters together with their spatial distribution and allow them to compare two lesions side by side. With our knowledge-assisted comparative visual analysis, physicians can explore and classify breast lesions. The true positive prediction accuracy of our scoring system achieved 71.4% in one-fold cross-validation using 14 lesions.