Development of efficient machine learning techniques for breast cancer diagnosis from histopathological nuclei Images

Abstract

Early detection of cancer by different screening methods has reduced the risk of newlinebreast cancer deaths in recent years. Analysis of tissue at the cellular level has shifted newlinefrom the microscope to the computer screen and glass slides have been converted to newlinedigital images for storage with digitalisation of the conventional pathology workflow. newlineComputer aided diagnosis (CAD) systems are integrated into the digital pathology newlineworkflow to assist pathologists for efficient and speedy diagnosis. newlineImplementation of digital pathology is not without challenges. High equipment newlinecost, high bandwidth required for faster transfer of data, substantial digital storage newlineneeded for high resolution images, high investment required for developing the newlineCAD system are some of the constraints faced in implementing digital pathology, newlineparticularly in low resource countries. This motivates us to develop an efficient, cost newlineeffective and computationally light CAD system for breast cancer diagnosis that newlinecan be incorporated into the digital pathology workflow. The main objective of the newlinemethods proposed in this work is to discriminate between benign and malignant newlinebreast tumors in histopathological images. newlineIn the first proposal, global feature engineering from preprocessed image is used newlinefor developing CAD architecture. To improve the capabilities of the handcrafted tex- newlineture features used, our method amended the traditional way of extracting segmented newlinefeatures by focusing primarily on the nuclei texture features globally. A variety newlineof state-of-the-art image descriptors and classifiers are tested in order to discover newlinethe best descriptor and classifier for distinguishing between benign and malignant newlinehistopathological images. An overall best sensitivity (true positive rate) of 95% and newlinean average image level accuracy of 90.51% are obtained when tested on the publicly newlineavailable BreaKHis dataset.To improve the nucleus guided feature engineering based CAD approach, the newlinefollowing techniques are introduced in the next proposal. newline