Hierarchical graph representations in digital pathology

ScoreNet: Learning Non-Uniform Attention and Augmentation for Transformer-Based Histopathological Image Classification

“…CNN's architectures tend to favor texture-based features [12], at the other end of the spectrum are graph neural network-based methods, which model the global dependencies between the local representations and thus rely more on the shape cues. We further compare with graph-based methods with a significant emphasis on HACT-Net [26], which holds the current state-of-art for TRoIs classification on the BRACS. ScoreNet reaches a new state-of-the-art weighted F1-score of 64.4% on the BRACS TRoIs classification task outperforming HACT- [22,29] 32.3 ± 4.6 39.0 ± 0.8 23.7 ± 1.7 18.0 ± 0.8 37.7 ± 2.9 47.3 ± 2.0 70.7 ± 0.5 39.4 ± 1.9 CNN (10× + 20×) [22,29] 48.3 ± 2.0 45.7 ± 0.5 41.7 ± 5.0 32.3 ± 0.9 46.3 ± 1.4 59.3 ± 2.0 85.7 ± 1.9 52.3 ± 1.9 CNN (10× + 20× + 40×) [22,29] [24] 58.8 ± 6.8 40.9 ± 3.0 46.8 ± 1.9 40.0 ± 3.6 63.7 ± 10.5 53.8 ± 3.9 81.1 ± 3.3 55.9 ± 1.0 CG-GNN [24] 63.6 ± 4.9 47.7 ± 2.9 39.4 ± 4.…”

“…To evaluate all samples, we per- form stratified 5-fold cross-validation. For HACT-Net, we use the available pre-trained weights and follow the code implementation of [26]. As HACT-Net sometimes fails to generate embedding, to have a fair comparison, we only evaluate those samples where HACT-Net could successfully produce embedding (around 95% of the BACH and 80% of CAMELYON16 dataset).…”

“…Datasets. The primary dataset is the BReAst Carcinoma Sub-typing (BRACS) dataset [26]. Experimental Setup.…”

“…Therefore their patch extraction is fixed and not datadriven as ours. To cope with that loss of information, graph neural network-based methods [26,39] have been proposed to consider global contextual information and the dependencies between the instances. These approaches build a graph model that operates on the cell-level structure or combines the cell-level and tissue-level context.…”

“…We extract two auxiliary datasets from the BRACS dataset [25]: a tile dataset at 40× and a low-resolution thumbnail dataset at 40 s ×, where s is the down-scaling ratio. The former dataset is used to pre-train the fine-grained attention module, whereas the latter serves to pre-train the recommendation stage.…”