Efficient automated detection of mitotic cells from breast histological images using deep convolution neutral network with wavelet decomposed patches

https://doi.org/10.1016/j.compbiomed.2018.11.001Get rights and content

Highlights

  • Proposed a new approach of design deep CNN architecture using wavelet decomposed patches.

  • Improved mitotic cell detection performances is achieved from breast cancer histological images.

  • Proposed approach is computationally efficient compared to the conventional deep CNN approach.

Abstract

In medical practice, the mitotic cell count from histological images acts as a proliferative marker for cancer diagnosis. Therefore, an accurate method for detecting mitotic cells in histological images is essential for cancer screening. Manual evaluation of clinically relevant image features that might reflect mitotic cells in histological images is time-consuming and error prone, due to the heterogeneous physical characteristics of mitotic cells. Computer-assisted automated detection of mitotic cells could overcome these limitations of manual analysis and act as a useful tool for pathologists to make cancer diagnoses efficiently and accurately. Here, we propose a new approach for mitotic cell detection in breast histological images that uses a deep convolution neural network (CNN) with wavelet decomposed image patches. In this approach, raw image patches of 81 × 81 pixels are decomposed to patches of 21 × 21 pixels using Haar wavelet and subsequently used in developing a deep CNN model for automated detection of mitotic cells. The decomposition step reduces convolution time for mitotic cell detection relative to the use of raw image patches in conventional CNN models. The proposed deep network was tested using the MITOS (ICPR2012) and MITOS-ATYPIA-14 breast cancer histological datasets and shown to outperform existing algorithms for mitotic cell detection. Overall, our method improves the performance and reduces the computational burden of conventional deep CNN approaches for mitotic cell detection.

Introduction

Among the various criteria for histopathological assessment of cancer, mitotic cell counting is one of the most important diagnostic approaches. Mitotic cells are produced when mother cells divide into two genetically identical cells that can differ in their morphological appearance. As a result, there is a general morphological heterogeneity in mitotic cells. Detection of cell proliferation markers is important in patient risk assessment for cancer. An increase in mitotic cell count signifies a higher proliferative index, which might reflect high cell proliferation activity that is observed in cancer. In conventional practice, experts manually determine mitotic nuclei count in images of tissue samples. Specifically, mitotic cell counts are conducted using a 400× microscopic magnification and reported as number of mitotic cells in an area of 2 mm2 [1]. It is broadly accepted that manual counting of mitotic cells is very error-prone due to the morphological and textural variations in microscopic images [2]. Therefore, high inter-observer variability is noted in manual mitotic cell counts.

Various computational approaches have been developed to overcome the drawbacks of manual mitotic cell counting. These computational approaches are thought to be superior because algorithms lack observer bias [3]. Different probabilistic models, such as the hidden Markov model [4], Gamma-Gaussian mixture model [5], maximum likelihood (ML) estimation [6,7], and Bayesian modeling [8], have been implemented for detecting mitotic cells in breast cancer histological images. Furthermore, a method has been proposed for mitotic cell segmentation that uses a distance regularized level set approach [9]. Another approach by Irshad et al. adopts a hierarchical model that uses scale-invariant feature transform to detect mitotic cells from breast histological images [10]. Supervised classification approaches, such as support vector machine [6,7,9,[11], [12], [13], [14]], cascade adaboosts [15], deep residual learning [16], and random forest [17], have also been used for mitotic cell identification. Combination of support vector machine and random forest have been used for the same mitotic cells detection. Unsupervised classification approach like k-nearest neighbors has been utilized for mitotic cell detection [18].

Deep convolution neural networks (CNNs) have been implemented widely in the medical image analysis domain [[19], [20], [21], [22]]. A CNN is a type of deep neural network with multiple hidden layers, and it has been used for mitotic cell detection in breast cancer histological samples [23]. Others have combined a cascade model of CNN with handcrafted, domain adaptive descriptors for mitotic cell segmentation as well [2]. Deep CNN is an attractive computational approach for mitotic cell counting due to its ability to comprehend and perform nonlinear mapping of input and output variables [24]. It is able to do so because it has millions of training parameters, which are optimized during the training procedure. However, the immense amount of hyperparameters that is used for mitotic cell counting in deep CCN models makes it highly complex computationally [[25], [26], [27], [28]]. Of interest, its time complexity is computed based on four parameters, namely patch size of the input image, number of channels, convolution kernel size, and the number of filters. Among them, the size of input patches contributes greatly to the computation complexity of the network. With a larger patch size, the complexity of the deep CNN increases. To reduce computational complexity, we decomposed input image patches with Haar wavelet and subsequently introduced these patches to the CNN architecture. Other wavelets, such as Daubechies (db)1, db2, db3, and db4, were tested, but the Haar wavelet turned out to be the most effective in decomposing input patches [29,30]. This approach of using Harr wavelet decomposed patches (21 × 21 pixels) with CNN outperformed conventional CNN models with raw patches (81 × 81 pixels) in mitotic cell detection of breast cancer histological images in both computational efficiency and accuracy. Overall, we propose an efficient, accurate CNN-based approach for automated mitotic cell detection that uses Harr wavelet decomposed patches to reduce computational burden.

Section snippets

Proposed mitotic cell segmentation framework

The proposed mitotic cell segmentation framework is shown in Fig. 1. It consists of data acquisition; stain normalization; candidate cell segmentation; patch creation; Haar wavelet decomposition of image patches; construction of a deep CNN architecture that learns the patterns of the decomposed data; mitotic cell detection in test datasets using our proposed CNN-based model and the conventional CNN model; and performance analysis of both CNN models.

Experimental results

In this study, we focused on identifying mitotic cells from breast histological images in an automated manner. We initially determined the most suitable input image patch size and convolution kernel or filter size for the conventional 18-layer CNN architecture (Fig. 7) to identify mitotic cells from raw images based on performance. Specifically, we tested the conventional CNN architecture with raw input patches of 61 × 61, 81 × 81, and 101 × 101 pixels and convolution kernels of 3 × 3 and 5 × 5

Discussion

The proposed method focused on improving the performance of deep CNN model-based automated mitotic cell detection systems by introducing a wavelet downsampling process on candidate mitotic cell image patches before subjecting them to a CNN architecture. Our method was applied on breast histological images. We demonstrated that our proposed CNN approach performed better than the conventional practice of using raw images patches for deep CNN. In particular, the patch image decomposition process

Conclusions

This study describes a new approach for mitotic cell detection in breast histological images. Specifically, a new CNN architecture that uses wavelet decomposed image patches is introduced here to reduce the computational complexity of conventional CNN models that use raw image patches. The proposed segmentation approach consists of candidate cell detection; patch extraction and their Haar wavelet decomposition; and mitotic cell detection using a CNN architecture that is design for decomposed

Conflicts of interest

Authors have no conflicts of interest to disclose.

Acknowledgements

The authors would like to thank Dr. Ludovic Roux of the IPAL Laboratory, A*STAR/I2R, Institute for Infocomm Research, CNRS UMI 2955, Singapore. Dr. Roux was an organizer of the MITOS-ATYPIA-14 grand challenge contest, and he independently evaluated the performance of our Harr-wavelet decomposed patch-based CNN approach on the test dataset of MITOS-ATYPIA-14. First author would like to acknowledge IIT Kharagpur for providing institute research fellowship grant to carry out this research.

References (46)

  • H. Wang et al.

    Cascaded ensemble of convolutional neural networks and handcrafted features for mitosis detection

    Proc. SPIE-Int. Soc. Opt. Eng.

    (2014)
  • A. Giusti et al.

    A comparison of algorithms and humans for mitosis detection

  • G.M. Gallardo et al.

    Mitotic cell recognition with hidden Markov models

    SPIE Int. Soc. Opt. Eng.

    (2004)
  • A.M. Khan et al.

    A gamma-Gaussian mixture model for detection of mitotic cells in breast cancer histopathology images

  • A. Tashk et al.

    A novel CAD system for mitosis detection using histopathology slide images

    J. Med. Signals Sens.

    (2014)
  • R. Nateghi et al.

    Automatic detection of mitosis cell in breast cancer histopathology images using genetic algorithm

  • C. Lu et al.

    Toward automatic mitotic cell detection and segmentation in multispectral histopathological images

    IEEE J. Biomed. Health Inform.

    (2014)
  • T. Wan et al.

    Wavelet-based statistical features for distinguishing mitotic and non-mitotic cells in breast cancer histopathology

    Proc. IEEE Int. Conf. Image Process.

    (2014)
  • C. Lu et al.

    Automated image analysis of nuclear atypia in high‐power field histopathological image

    J. Microsc.

    (2015)
  • C. Sommer et al.

    Learning-based mitotic cell detection in histopathological images

  • R. Nateghi et al.

    Maximized inter-class weighted mean for fast and accurate mitosis cells detection in breast cancer histopathology images

    J. Med. Syst.

    (2017)
  • T. Wollmann et al.

    Deep residual Hough voting for mitotic cell detection in histopathology images

  • A. Paul et al.

    Enhanced random forest for mitosis detection

  • Cited by (27)

    • Computer assisted mitotic figure detection in histopathology images based on DenseNetPCA framework

      2022, Materials Today: Proceedings
      Citation Excerpt :

      Their model resulted in F-score of 0.88 and shows improved accuracy. Das DK et al.,[6] applied deep CNN along with wavelet decomposed patches. The study reveals that wavelet decomposition of image patches helps to reduce the time complexity of the deep network.

    • Breast cancer intelligent analysis of histopathological data: A systematic review

      2021, Applied Soft Computing
      Citation Excerpt :

      The difficulty in analyzing mitoses in WSI is similar to analyzing cell nuclei in WSI of breast cancer. In this way, the analysis of mitosis is usually carried out at the patch level [40,44,45]. Another way to decrease the complexity of atypical mitoses analysis is to perform cell nuclei segmentation [46].

    • Automatic discrimination of neoplastic epithelium and stromal response in breast carcinoma

      2019, Computers in Biology and Medicine
      Citation Excerpt :

      In the last decade, several algorithms have been proposed for the automatic analysis of breast carcinoma [8–10]. These methods are based on multi-resolution algorithms [11,12], nuclear features [13,14], and deep neural networks [15–17] to perform cancer detection and grading. In particular, deep neural networks have driven advances in image recognition and they achieved state-of-art performance in many segmentation tasks of medical imaging.

    • How to use AI in pathology

      2023, Genes Chromosomes and Cancer
    View all citing articles on Scopus
    View full text