Correlation Matters: Multi-scale Fine-Grained Contextual Information Extraction for Hepatic Tumor Segmentation

Automatic tumor segmentation has been used as a diagnostic aid in the identification of diseases such as tumors from liver CT scans, and their treatment. Owing to their success in computer vision tasks, the state-of-the-art Fully Convolutional Networks (FCNs) or U-Net based models have often been employed in many recent studies for automatic tumor segmentation to learn numerous weight-shared convolutional kernels and extract various semantic features. However, the correlation between different tumor regions in feature maps cannot be easily captured due to the lack of contextual dependencies, which in turn limits the representative capability of the adopted models and thus affects the accuracy of tumor segmentation results. To resolve this issue, we propose a novel framework for segmentation of tumors in liver CT scans, which can explicitly extract multi-scale fine-grained contextual information by adaptively aggregating local features with their global dependencies. The proposed multi-scale framework features a light model with a very few additional parameters, and also its visualization capability significantly boosts networks’ interpretability. Experimental results on a real-world liver tumor CT dataset illustrate that the proposed framework achieves the state-of-the-art performance in terms of a number of widely used evaluation criteria for the hepatic tumor segmentation task.