MAQLIN PARAMANANDAM et al.: AUTOMATED DETECTION OF MITOTIC FIGURES IN BREAST CANCER HISTOPATHOLOGY IMAGES USING GABOR FEATURES AND DEEP NEURAL NETWORKS 1366 AUTOMATED DETECTION OF MITOTIC FIGURES IN BREAST CANCER HISTOPATHOLOGY IMAGES USING GABOR FEATURES AND DEEP NEURAL NETWORKS Maqlin Paramanandam 1 , Robinson Thamburaj 2 and Joy John Mammen 3 1,2 Department of Mathematics, Madras Christian College, India E-mail: 1 maqlinparamanandam@yahoo.com, 2 robinson@mcc.eud.in 3 Department of Transfusion Medicine, Christian Medical College Hospital, India E-mail: 3 joymammen@cmcvellore.ac.in Abstract The count of mitotic figures in Breast cancer histopathology slides is the most significant independent prognostic factor enabling determination of the proliferative activity of the tumor. In spite of the strict protocols followed, the mitotic counting activity suffers from subjectivity and considerable amount of observer variability despite being a laborious task. Interest in automated detection of mitotic figures has been rekindled with the advent of Whole Slide Scanners. Subsequently mitotic detection grand challenge contests have been held in recent years and several research methodologies developed by their participants. This paper proposes an efficient mitotic detection methodology for Hematoxylin and Eosin stained Breast cancer Histopathology Images using Gabor features and a Deep Belief Network- Deep Neural Network architecture (DBN-DNN). The proposed method has been evaluated on breast histopathology images from the publicly available dataset from MITOS contest held at the ICPR 2012 conference. It contains 226 mitoses annotated on 35 HPFs by several pathologists and 15 testing HPFs, yielding an F-measure of 0.74. In addition the said methodology was also tested on 3 slides from the MITOSIS- ATYPIA grand challenge held at the ICPR 2014 conference, an extension of MITOS containing 749 mitoses annotated on 1200 HPFs, by pathologists worldwide. This study has employed 3 slides (294 HPFs) from the MITOS-ATYPIA training dataset in its evaluation and the results showed F-measures 0.65, 0.72and 0.74 for each slide. The proposed method is fast and computationally simple yet its accuracy and specificity is comparable to the best winning methods of the aforementioned grand challenges. Keywords: Breast Cancer, Deep Neural Networks, Gabor Filter, Histopathology, Mitotic Count 1. INTRODUCTION In Breast cancer pathology, the single factor that best aids in establishing the proliferative activity of the tumor is the number of cells undergoing mitotic division visible under a fixed number of high power fields (HPF - the area of tissue under the microscope set to a high magnification). Studies reveal that mitotic count is considered the most independent prognostic parameter that determines patient risk [1] and is assessed through the strictest of protocols. However, mitotic counting is highly subjective, prone to inter and intra observer variability [2]. An automated detection of mitotic figure using image analysis could be an efficient, error free and time saving also making results obtained by different pathologists comparable. Automated mitotic detection has certain innate challenges due to the high-complexity in appearance. The most prominent feature of a cell undergoing mitotic division is its hyperchromaticity, and effective care needs to be taken to avoid counting other hyperchromatic elements such as lymphocytes or apoptic nuclei as mitoses. Another challenge is the variability in the shapes of mitosis in its four main phases: prophase, metaphase, anaphase and telophase shown in Fig.1(a-d) respectively. Specifically, a mitotic cell in telophase, though having two separate and fully divided nuclei, should be counted as a single mitotic figure. The Fig.1(e)-Fig.1(h) shows certain hyperchromatic nuclei which have close resemblance to mitotic figures. (a) (b) (c) (d) (e) (f) (g) (h) Fig.1. (a-d) Different phases of mitotic figures, (e-h) Other Hyperchromatic nuclei in the images which closely resemble mitotic figures Given the significance of the Mitotic count and the related issues, this paper proposes an efficient mitotic detection methodology for H&E stained breast cancer histopathology images. From input breast cancer histopathology image the proposed method detects mitotic figures using automated image analysis and a trained Deep Belief Network - Deep Neural Network classifier (DBN-DNN). The method is trained and evaluated on breast histopathology images from dataset presented at MITOS contest held at ICPR 2012 conference [3] and MITOSIS- ATYPIA grand challenge held at ICPR 2014 [4]. The paper is organized as follows: Section 2 gives a short review of the related works. Section 3 describes the dataset and ground truth followed by a presentation of methodology in Section 4 and the results and concluding remarks are laid out in Section 5 and Section 6 respectively. 2. RELATED WORKS Very little documentation on automated mitotic detection can be found, on research done two decades ago, owing to the fact that only limited computational and tissue digitization resources were available at the time. Recent interest was kindled after the advent and widespread usage of Whole Slide scanners [5], [6]. To help DOI: 10.21917/ijivp.2016.0199