Classification of Mammographic Microcalcification Clusters

  • Nashid Alam

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

Death rates for women suffering from breast cancer are high. This is often because of delayed diagnosis, which occurs at a stage at which the cancer is advanced and therefore impervious to treatment. Early detection of breast abnormalities is crucial for increasing cancer survival rates. Microcalcifications (MC) — small flecks of calcium salts — can be found in routine mammogram screenings. Their presence in specific patterns or clusters can be a sign of precancerous cells or early breast cancer. Radiologists use human perception to identify MC features, and this can result in perception error which can lead to late diagnosis. This often happens because Microcalcifications (MC) clusters are difficult to perceive as they tend to be very small in size, have irregular shapes and are sometimes located in dense breast tissue with low visibility. There is an unmet need for Computer-aided Diagnosis (CADx) tools that can provide a “second opinion” to help radiologists with their diagnoses. Towards this goal, this thesis proposes automatic frameworks for classifying and localising MC clusters using mammograms. This includes automatic segmentation of MC clusters and identification of benign and malignant clusters based on their sizes, shapes, textures, and spatial distributions. Our project proceeds as follows. Firstly, MC cluster segmentation methods are proposed by incorporating image interpolation, morphological operations and edge-preserving filtering techniques. The segmentation results are compared with radiologist's annotation using Dice Similarity Score (DSC). Spatial distribution features extracted from the segmented images outperformed when compared with the size, shape and texture features to classify MC clusters for Digital Database for Screening Mammography (DDSM), Mammographic Image Analysis Society (MIAS) and Optimam Mammography Image Database (OMI-DB) datasets. Annotations in these datasets include associated biopsy results which are cross-checked with the results obtained from our MC cluster classifier. A stack-generalisation classifier is developed to classify MC clusters which provided 96.47% Classification Accuracy (CA). Secondly, a DL based model is proposed for MC clusters classification and localisation using a modified pre-trained classifier with a minimal training data using the OMI-DB datasets. Our proposed DL model provided 94.12% Classification Accuracy (CA). When comparing the result with stack-generalisation it is revealed that for some specific task, the more traditional methods performs better than the DL model, which is inline with the limitations of DL presented in the literature review. Finally, We adopt the 3L algorithm for Implicit B-spline (IBS) fitting to investigate the robustness of Three-dimensional (3D) models of MC clusters to classify benign and malignant cases. Features were extracted from the 3D models of MC clusters. In this preliminary study, we obtained an 80.00% CA. We analysed images from five women from an anonymised Digital Breast Tomosynthesis (DBT) dataset (acquired at the University of Pennsylvania Breast Imaging Center). The ground truth of the annotation of this dataset is based on biopsy results.
Date of Award2023
Original languageEnglish
Awarding Institution
  • Aberystwyth University
SupervisorReyer Zwiggelaar (Supervisor) & Bernie Tiddeman (Supervisor)

Cite this

'