This innovation utilises sparse representation classification with dictionary learning to classify normal/abnormal cells. It also facilitates compression of pathology slides.
Histopathological image analysis has some unique challenges to be addressed. Very large image sizes and insufficient labelled images are two of the major issues. This innovation proposes a solution to find optimised representation domain which facilitates both classification and compression of pathology slide images. The proposed system learns efficient class-specific dictionaries from the existing labelled histopathological images. Obtained dictionaries are then used for classification of “normal” or “abnormal” test samples. Our unique novel method works based upon projective dictionary pair learning which has already shown promising performance in classification of face and handwritten images. We will further expand our method to deep dictionary classification to address multiple tissue slice nature of pathology images. The proposed system is superior to the existing deep representations trained for natural images and not suitable for pathology images. Unlike deep learning networks requiring huge amount of labelled data, the proposed method requires less labelled data for training. In addition, since the obtained dictionaries allow sparse data representation, they ultimately facilitate data compression.
Improving analysis, annotation, and storage of digital images
Achieving higher classification accuracy of digital images
Reducing the number of meta-parameters compared to existing approaches