Machine Learning-Aided Automatic Detection of Breast Cancer: A Survey

Introduction

Deep Learning approach has gained tremendous recognition in the area of computer vision particularly in object detection, image segmentation and recognition (Mugahed A. Al-antari T.-S. K., 2020). Precise and premature detection and classification of lesions is highly critical for increasing the survival rates of persons affected. According to the statistics published by Indian Council of Medical Research (ICMR), India is likely to have over 17.3 lakh new cases of cancer causing over 8.8 lakh deaths by 2020 with cancers of breast, lung and cervix topping the list. Breast Cancer with an estimate of 8.8 lakh new cases in 2020 is likely to outnumber any other type of cancer in India. The prognosis of such types of disease is a challenging task for radiologists due to the huge number of cases together with careful and thorough examination it demands. Mammography is the most extensively used procedure for breast cancer screening and has led to remarkable reduction of death rate through prior detection. However, the complexity of mammograms, histopathological slides, MRIs and other type of scans together with massive volume of inspections per radiologist can result in false diagnosis. Computer-aided detection (CAD), employing image processing and pattern recognition techniques should provide a second opinion assisting radiologists for cancer diagnosis. Recent studies confirm CAD models cannot further significantly improve the diagnostic accuracy of mammograms in particular (Lehman CD, 2015). The main hindrance in utilizing CAD for lesion detection is high false positive rates (FPR). Moreover, there exists a trade-off between sensitivity and specificity when primitive CAD models are used. The consequence is increased patient anxiety, auxiliary radiation exposures, needless biopsies and unnecessary high health care costs. The constraints of present-day CAD necessitate new and accurate detection procedures, because millions of people undergo various types of scans each year and even a modest decrease in the FPR is an immense gain.

With the phenomenal achievement in machine learning especially deep learning, with respect to visual object recognition and detection, there is an ample scope in establishing deep learning tools to aid radiologists and enhance the accuracy of screening mammograms or histopathological slides. Contemporary research has demonstrated that deep learning assisted CAD models perform very well and also improve the radiologist`s decision making (Lee JG, 2017). In comparison to autonomous CAD systems, Convolutional Neural Networks (CNNs) attain higher accuracy results and provide quantitative analysis of suspicious lesions (Kooi T, 2016). Wang et al. (Wang J, 2017) have reported that Deep Learning approaches drop human error rate for cancer diagnosis by 85%.

Detection of subclinical tumour on screening mammograms is a challenging task in the classification of lesions because these lesions are immensely variable in terms of location, shape, size and texture. For example, a mammogram image is generally 4000 x 3000 px, at the same time a probable region of interest (ROI) may be as small as 100 x 100 px. This gives a good reason for many researches to focus on classification of annotated lesions. Despite the fact that classification of ROIs is a preliminary step, an automated model must process the entire input image to supplement the known information and enhance clinical interpretations. If ROIs were extensively available in medical databases, then traditional detection and classification approaches could be easily tested. However, the methods that need ROI annotations cannot be transferred to sizable mammographic databases that are scarce of ROI annotations. Proper identification of probable lesions is very essential in obtaining a high true positive rate (TPR), enhancing the diagnosis of tumours. The traditional techniques rely on less specific and hand-crafted features that lack the precision in the task of automatic detection of probable lesions. The present-day convolutional neural network (CNN) models are designed to enhance radiologist`s understanding of the identification of even the least possible lesions at a very early stage. In this article we review the latest contributions to deep learning domain in detecting and classifying lesions using various medical images.