Abstract
Histopathology image is an important basis for pathologists to evaluate disease at the cellular level and colon cancer tissue sections usually contain many different types of nuclei, which should be automatically detected and identified. However, the detection and the identification of cell nuclei are challenging tasks due to the complex tissue structure and the diversity of nuclear morphology. In this paper, a staged detection-identification framework is proposed for cell nuclei in colon cancer histopathology images. First, nuclei positions are detected by a position of interest network, which encodes context-aware representation on input image and decodes features on proximity map. Meanwhile, a cascade residual fusion block is presented to enhance the detection performance during the decoding process. Second, a multicropping network is developed to identify the detected cell nuclei. For reducing the impact of uncertainty, a multicropping module is designed for effectively capturing contextual feature contents around the center of a nucleus. The proposed detection-identification framework is evaluated on an available colorectal adenocarcinoma images data set, which has 100 images including more than 20 000 marked nuclei. Compared with state-of-the-art methods, the proposed approach demonstrates excellent performance with better prediction scores.
Original language | English |
---|---|
Article number | 8640829 |
Pages (from-to) | 183-193 |
Number of pages | 11 |
Journal | IEEE Transactions on Instrumentation and Measurement |
Volume | 69 |
Issue number | 1 |
DOIs | |
Publication status | Published - Jan 2020 |
Keywords
- Deep convolutional neural network
- histopathology image
- nucleus detection
- pattern recognition