Micro and Macro Breast Histology Image Analysis by Partial Network Re-use

摘要

Convolutional neural networks (CNN) have shown to be effective in medical image processing and analysis. Herein, we propose a CNN approach to perform patch- and pixel-wise histology labeling on breast microscopy and whole slide images (WSI), respectively. We devise a processing block that is capable of extracting compact features in an efficient manner. Based upon the processing block, classification and segmentation networks are built. Two networks share an encoder via partial transformation and transfer learning to maximally utilize the trained network and available dataset. 400 microscopy images and 10 WSI were employed to evaluate the proposed approach. For patch classification, an accuracy of 71% and 65% were obtained on the training and testing dataset, respectively. As for segmentation, we achieved an overall score of 0.7343 and 0.4945 on the training and testing dataset, respectively.