Table 2 Macro F-score averaged over each tissue type for the SA-HCNN with PTO tiling, the SA-HCNN with Center tiling, the current best baseline CNN (EfficientNetV2) with PTO tiling, SegFormer with center tiling, and Segformer with PTO tiling. Results show that the SA-HCNN approach significantly outperforms the baseline EfficientNetV2 and image segmentation approach, and the PTO tiling provides a small additional performance improvement over Center tiling.

From: Scalable deep learning artificial intelligence histopathology slide analysis and validation