Ensembling Neural Networks for Digital Pathology Images Classification and Segmentation

Abstract

In the last years, neural networks have proven to be a powerful framework for various image analysis problems. However, some application domains have specific limitations. Notably, digital pathology is an example of such fields due to tremendous image sizes and quite limited number of training examples available. In this paper, we adopt state-of-the-art convolutional neural networks (CNN) architectures for digital pathology images analysis. We propose to classify image patches to increase effective sample size and then to apply an ensembling technique to build prediction for the original images. To validate the developed approaches, we conducted experiments with Breast Cancer Histology Challenge dataset and obtained 90% accuracy for the 4-class tissue classification task.

A New Ensemble Learning Framework for 3D Biomedical Image Segmentation

“…Since the results from the base-learners can be very close to the ground truth, training the meta-learner by directly fitting y i is susceptible to over-fitting. Many measures have been proposed to address this issue: (1) simple meta-learner structure designs (e.g., in (Ju, Bibaut, and van der Laan 2018), the metalearner was implemented as a single 1 × 1 convolution layer, and in (Makarchuk et al 2018), the meta-learner was implemented using the XGBoost classifier (Chen and Guestrin 2016)); (2) excluding raw image information from the input (Zhou 2012) However, in our 3D biomedical image segmentation scenario, these meta-learner designs may not work well due to the following reasons. First, each of our individual baselearners (2D and 3D models) has its distinct merit; in many difficult image areas, it is quite likely that only one of the base-learners could produce the correct results.…”

“…However, recent studies have not leveraged this property to design a better meta-learner for image segmentation. For example, in (Ju, Bibaut, and van der Laan 2018;Nigam, Huang, and Ramanan 2018), only linear combination of base-learners was explored. We develop a new fully convolutional network (FCN) based meta-learner to capture the merits of our base-learners and produce spatially consistent results.…”