Structure-aware scale-adaptive networks for cancer segmentation in whole-slide images
This addresses cancer assessment for medical imaging by improving segmentation of vague boundaries and small regions, though it appears incremental as it builds on existing encoder-decoder architectures with novel modules.
The paper tackled cancer segmentation in whole-slide images by proposing a structure-aware scale-adaptive feature selection method, achieving outstanding performance on liver cancer segmentation compared to top results in the PAIP 2019 challenge and showing improvements in colorectal cancer segmentation.
Cancer segmentation in whole-slide images is a fundamental step for viable tumour burden estimation, which is of great value for cancer assessment. However, factors like vague boundaries or small regions dissociated from viable tumour areas make it a challenging task. Considering the usefulness of multi-scale features in various vision-related tasks, we present a structure-aware scale-adaptive feature selection method for efficient and accurate cancer segmentation. Based on a segmentation network with a popular encoder-decoder architecture, a scale-adaptive module is proposed for selecting more robust features to represent the vague, non-rigid boundaries. Furthermore, a structural similarity metric is proposed for better tissue structure awareness to deal with small region segmentation. In addition, advanced designs including several attention mechanisms and the selective-kernel convolutions are applied to the baseline network for comparative study purposes. Extensive experimental results show that the proposed structure-aware scale-adaptive networks achieve outstanding performance on liver cancer segmentation when compared to top ten submitted results in the challenge of PAIP 2019. Further evaluation on colorectal cancer segmentation shows that the scale-adaptive module improves the baseline network or outperforms the other excellent designs of attention mechanisms when considering the tradeoff between efficiency and accuracy.