2022/02/15 by Thomas Stegmüller, Behzad Bozorgtabar, Stegmüller, Thomas +5 · 1 citation
Computer Science · Engineering · #AI in cancer detection #Digital Imaging for Blood Diseases #Medical Imaging and Analysis
paper · pdf · doi:10.48550/arxiv.2202.07570
Progress in digital pathology is hindered by high-resolution images and the\nprohibitive cost of exhaustive localized annotations. The commonly used\nparadigm to categorize pathology images is patch-based processing, which often\nincorporates multiple instance learning (MIL) to aggregate local patch-level\nrepresentations yielding image-level prediction. Nonetheless, diagnostically\nrelevant regions may only take a small fraction of the whole tissue, and\ncurrent MIL-based approaches often process images uniformly, discarding the\ninter-patches interactions. To alleviate these issues, we propose ScoreNet, a\nnew efficient transformer that exploits a differentiable recommendation stage\nto extract discriminative image regions and dedicate computational resources\naccordingly. The proposed transformer leverages the local and global attention\nof a few dynamically recommended high-resolution regions at an efficient\ncomputational cost. We further introduce a novel mixing data-augmentation,\nnamely ScoreMix, by leveraging the image's semantic distribution to guide the\ndata mixing and produce coherent sample-label pairs. ScoreMix is embarrassingly\nsimple and mitigates the pitfalls of previous augmentations, which assume a\nuniform semantic distribution and risk mislabeling the samples. Thorough\nexperiments and ablation studies on three breast cancer histology datasets of\nHaematoxylin & Eosin (H&E) have validated the superiority of our approach over\nprior arts, including transformer-based models on tumour regions-of-interest\n(TRoIs) classification. ScoreNet equipped with proposed ScoreMix augmentation\ndemonstrates better generalization capabilities and achieves new\nstate-of-the-art (SOTA) results with only 50% of the data compared to other\nmixing augmentation variants. Finally, ScoreNet yields high efficacy and\noutperforms SOTA efficient transformers, namely TransPath and SwinTransformer.\n