A multi-scale 3-stacked-layer coned U-net framework for tumor segmentation in whole slide images

The contribution of deep learning in medical image diagnosis has gained extensive interest due to its excellent performance. Furthermore, the interest has also grown in digital pathology since it is considered the golden standard for tumor detection and diagnosis in digital Whole Slide Images (WSIs). This paper proposes an end-toend cone-shaped encoder-decoder framework called a Multi-scale 3-stacked-Layer coned U-Net (Ms3LcU-Net) framework. It boosts performance by using many enhancements and integrating techniques such as blended mutual attention, dilated fusion, edge enhancement, and atrous pooling. Furthermore, the morphological postprocessing and test time augmentation techniques are used in Ms3LcU-Net to refine and smooth the generated segmentations. The experimental results from a quantitative perspective using multiple evaluation metrics and from a qualitative viewpoint by visualizing the generated segmentation predictions conducted on the public PAIP 2019 and DigestPath datasets demonstrated the effectiveness and competitiveness of the proposed model for tumor segmentation in WSIs. The proposed framework yielded an average clipped Jaccard Index value of 0.7211 on the validation set of the PAIP 2019 dataset. In contrast, the DigestPath dataset achieved an average dice coefficient and F1-score of 0.833 and 0.897, respectively. The code will be available publicly upon acceptance of the paper at https://github.com/Heba-AbdeNabi/Ms3LcU-Net-.