A hard segmentation network guided by soft segmentation for tumor segmentation on PET/CT images

Cancer is considered one of the leading causes of death. We can detect cancers through the anatomical and functional imaging provided by PET/CT. However, many tumors in PET/CT are obvious in only one modality, and PET contains many non-lesional hypermetabolic regions, which increases the difficulty of segmentation. Furthermore, traditional two-stage segmentation improves segmentation efficiency by breaking down a segmentation task into two independent subtasks. The second stage loses most of the feature information obtained in the first stage. To address these problems, we propose a hard segmentation network guided by soft segmentation for tumor segmentation on PET/CT images. The proposed network has a soft segmentation branch and a hard segmentation branch. The output of the soft segmentation branch is a logits map composed of gradient values, which is corrected with the soft ground truth by the proposed similarity loss function so that the logits map and the soft ground truth are approximately consistent in the high-dimensional vector space. The output of the hard segmentation branch is the final prediction map. The two branches are connected by a soft segmentation-guided mechanism. This guidance mechanism can generate a soft segmentation-guided map with stable distribution according to the logits map obtained by the soft segmentation branch. We validated the proposed network on two datasets. The Dice of 0.7324 on the public soft tissue sarcoma dataset and 0.7693 on the private liver tumor dataset. By only using U-Net as the backbone network, our method achieves the best performance.