A Novel Ensemble Strategy With Enhanced Cross Attention Encoder-Decoder Framework for Tumor Segmentation in Whole Slide Images

Due to the rapid increase in the quality of the generated medical images combined with the development of computational algorithms, Computer-Aided Diagnostic (CAD) systems, which are empowered with artificial intelligence to analyze these images, provide an objective decision to assist and guide the medical staff. In cancer diagnosis and staging, the golden standard is the microscopic examination of suspicious tissue either in the laboratory using the conventional glass slides or using their digital scanned counterparts called Whole Slide Images (WSI). However, the vast search space in these slides and the existing heterogeneous patterns in the sampled tissue make their analysis process labor-intensive and time-consuming. Therefore, this paper proposed a novel encoder-decoder framework with a new ensemble strategy to address the difficulties in semantically segmenting tumors in WSI. We introduced new enhancements to the base U-Net, including cross attention to aggregate information from different portions of the network. Moreover, we manipulated each stained slide to produce three new slides with different characteristics and fed them to separate models and injected their predictions together at the inference stage through our proposed ensemble strategy to form the final segmentation prediction. We tested and validated the proposed cross attention U-Net model with the ensemble strategy on the task of viable tumor segmentation of the challenging hepatocellular carcinoma in liver WSIs in the PAIP2019 dataset. The experimental results have demonstrated that the proposed ensemble strategy outperformed the single conventional handling of the WSI and improved the segmentation accuracy.