Study On Breast Histopatholgy Image Analysis With Deep Learning Based Approaches

Nowadays Digital pathology is one of the challenging evolutions in medical protocols.Pathological exams play a critical role in the diagnosis process and enable pathologists to categorize the microscopic structures.The pathologist analyzes numerous biopsy slides under the microscope.The analysis of histological structure of cell nuclei,morphology variations and tissue distributions of organisms assist the pathologists to better recognize histopathological samples.Histopathological classification and grading of high-content biopsy provide important prognostic information that is essential to know the disease(cancer)proliferation and prognosis.Among the predominant cancers,breast cancer is one of the main causes of cancer deaths impacting women worldwide,increasing gradually in developed and developing countries.However,manually diagnosing a disease in whole slide images(WSIs)of breast tissue is a laborious and challenging task.Various automatic diagnosis systems(i.e.,simple image processing algorithms or deep learning-based algorithms)have been developed to overcome the drawbacks of manual analysis.In this study,the main focus is on developing deep learning based automat detection frameworks,able to provide breast cancer detection from different types of histopathology images.World Health Organization recommended a standard for breast cancer grading called Nottingham grading system.It combines three morphological prognostic factors namely,mitotic count,tubule formation,and nuclear pleomorphism.In this study,theoretical guidance about histopathological image analysis is given,which improves the diagnostic and prognostic capabilities of histopathologists.At the same time,a comprehensive review of state-of-the-art approaches for mitotic cell detection,segmentation and histopathology image inference is presented.Histopathology image analysis poses a difficult computer vision problem,due to high variability in routinely hematoxylin and eosin(H&E)stained images.These varying conditions in slide preparation and morphological variations in tissue structures induce chances of misdiagnosis.Therefore,state-of-the-art color normalization methods are studied in detail,which can be helpful to reduce color variations in histology images.A robust stain color normalization approach has been designed and its performance is analyzed on different types of histology datasets.The deep learning methods particularly convolutional neural networks(CNN)have emerged as the prevailing paradigm for computational pathology problems.For all of the histology tasks,deep learning is able to outperform manual detection methods.In cancer diagnosis and grading,various automatic detection frameworks have been designed for different types of microscope datasets.In this dissertation four CNN based mitosis detection frameworks,fully fused convolution neural network(FF-CNN),Multi scale fully fused convolution neural network(MFF-CNN),Multi scale region proposal convolution neural network(MS-RCNN)and Atrous fully convolutional neural network(A-FCNN)frameworks are proposed for mitosis detection.One haar wavelet decomposed image based convolution neural network(HWDCNN)framework for multi-class classification of breast histology dataset is also designed.The FF-CNN model combines rich features from different level layers for mitosis discrimination.In this model,the concept of frequency weighted loss function solves the class imbalance problem in the training dataset.An up-sampling approach is implemented,which considerably reduces the computations and ensures a comparative accuracy at same time.Further,the domain agnostic deep MFF-CNN model performs mitoses detection by fusing multi-level and multi-scale features rather than single-level features from the last layer.The intended model,explore multistep fine-tuning phenomena and implements a multi-scale frequency weighted loss function to reduce network over-fitting.Since exhaustive pixel wise annotation is time consuming process so in most of breast histology mitosis datasets e.g.,ICPR 2014 and AMIDA13 datasets,pathologists provide annotations where only centroid pixel is labeled.To address this issue,a new atrous convolution based annotation(A-FCN)technique is designed that could be helpful for a pathologist to exhaustively annotate the mitotic cells from high-power field(HPF)images with perfect consensus.Our proposed A-FCN approach is also a promising solution to train a bounding box region proposal detector in a weakly supervised fashion.Moreover,mitosis detection is considered as object detection problem and detection of too small mitotic cells is addressed for the first time.The designed MS-RCNN detector incorporates several strategies that increase its performance on small size mitosis and thereby boost detection performance.Finally,a practical framework referred to as HWDCNN for multiclass recognition in breast histology images is proposed.In the design model,a concept of high resolution histopathological image decomposition with wavelet transform is introduced.Decomposition of images considerably reduces convolution time and computational resources of the deep CNNs,without any performance downgrade.Deep CNN based models usually demand lots of data,however,limited amount of available standard histology dataset prohibits the generalization power of these deep models.The demand for massive histology dataset is solved by virtue of transfer learning and a novel data augmentation approach.The data augmentation method generates broad range of realistic H&E stained images by perturbing stain intensity in each Hematoxylin and Eison color channels of original microscopy images.From a more global standpoint,our frameworks are promising solutions to breast cancer detection applications.