Breast Cancer Histological Image Classification Using Fine-Tuned Deep Network Fusion
Amirreza Mahbod, Isabella Ellinger, Rupert Ecker, Örjan Smedby, and Chunliang Wang
Springer International Publishing AG, part of Springer Nature 2018 A. Campilho et al. (Eds.): ICIAR 2018, LNCS 10882, pp. 754–762, 2018. doi.org/10.1007/978-3-319-93000-8_85
Histopathological image analysis, which is used for recognition and diagnosis of tissue abnormalities such as malignant lesions, is commonly performed by experienced pathologists. Computer-aided diagnosis (CAD) systems are semi or fully automated image analysis algorithms, which are developed to help pathologists during the diagnosis procedure. Being a second opinion system, CAD systems are supposed to reduce the workload of specialists, to improve the diagnosis efficiency, to increase the level of inter-observer agreement and, in the end, also contribute to cost reduction.
While there are a large number of automatic or semi-automatic methods for image analysis, machine learning-based algorithms have shown to be superior over conventional image processing techniques when multiple pattern types have to be recognized and distinguished. Convolutional Neural Networks (CNNs) are machine learning approaches for image analysis that do not rely on hand-crafted features, but utilize large amount of images to derive task-specific image features. While extremely promising, the research community still has to develop – and validate – appropriate algorithms.
Amirreza Mahbod is PhD student in an EU-funded European Training Network (ETN; Grant #675228, coordinated by Enikö Kallay (IPA); https://casr.meduniwien.ac.at) and co-supervised by Isabella Ellinger (IPA) and Rupert Ecker (CEO TissueGnostics GmbH). Amirreza makes use of deep CNNs for segmentation and classification of histopathological images with minimum pre- and post-processing steps. He utilizes state-of-the-art deep learning-based architectures and adapts them for histopathological image analysis. The algorithms are tested and validated in „Grand Challenges in biomedical image analysis“ such as the BACH (ICIAR 2018 Grand Challenge on BreAst Cancer Histology Images). The goal of the BACH challenge was to automatically classify H&E stained breast histology microscopy images in four classes: normal, benign, in situ carcinoma and invasive carcinoma. Amirreza´s method, which is described in the presented conference paper, was ranked place 18 of 51 participants. The performance of the method was evaluated based on the overall prediction accuracy. The ultimate goal is to integrate successfull classification and segmentation methods into the StrataQuestTM software of TissueGnostics GmbH and make them available for clinical as well as biomedical research applications.