An investigation of automatic feature extraction for clustered microcalcifications on digital mammograms

Abstract

Mammography is a common imaging modality used for breast screening. The limitations in reading mammogram images manually by radiologists have motivated an interest to the use of computerised systems to aid the process. Computer-aided diagnosis (CAD) systems have been widely used to assist radiologists in making decision; either for detection, CADe, or for diagnosis, CADx, of the anomalies in mammograms. This thesis aims to improve the sensitivity of the CADx system by proposing novel feature extraction techniques. Previous works have shown that multiple resolution images provide useful information for classification. The wavelet transform is one of the techniques that is commonly used to produce multiple resolution images, and is used to extract features from the produced sub-images for classification of microcalcification clusters in mammograms. However, the fixed directionality produced by the transform limit the opportunity to extract further useful features that may contain information associated with the malignancy of the clusters. This has driven the thesis to experiment on multiple orientation and multiple resolution images for providing features for microcalcification classification purposes. Extensive and original experiments are conducted to seek whether the multiple orientation and multiple resolution analysis of microcalcification clusters features are useful for classification. Results show that the proposed method achieves an accuracy of 78.3%, and outperforms the conventional wavelet transform, which achieves an accuracy of 64.9%. A feature selection step using Principal Component Analysis (PCA) is employed to reduce the number of the features as well as the complexity of the system. The overall result shows that the accuracy of the system when 2-features from steerable pyramid filtering are used as input achieved 85.5% as opposed to 2-features from conventional wavelet transform, which achieves an accuracy of 69.9%. In addition, the effectiveness of the diagnosis system also depends on the classifier. Deep belief networks have demonstrated to be able to extract high-level of input representations. The ability of greedy learning in deep networks provide a highly non-linear mapping of the input and the output. The advantage of DBN in being able to analyse complex patterns, in this thesis, is exploited for classification of microcalcification clusters into benign or malignant sets. An extensive research experiment is conducted to use DBN in extracting features for microcalcification classification. The experiment of using DBN solely as a feature extractor and classifier of raw pixel microcalcification images shows no significant improvement. Therefore, a novel technique using filtered images is proposed, so that a DBN will extract features from the filtered images. The analysis result shows an improvement in accuracy from 47.9% to 60.8% when the technique is applied. With these new findings, it may contribute to the identification of the microcalcification clusters in mammograms.