DigestPath: A benchmark dataset with challenge review for the pathological detection and segmentation of digestive-system

MS-CLAM: Mixed supervision for the classification and localization of tumors in Whole Slide Images

“…On the latter, using only 62% of the tile-level labeled slides would have been enough to reach the 1 st position on the challenge leaderboard based on the AUC results (5 th position on the final leaderboard). On DigestPath2019 on the other hand, using 6% of the tile-level labeled slides would have reached the 5 th rank in terms of AUC on the second task of the challenge (Da et al, 2022). For both of the challenges, the best results were obtained with the help of deep neural networks trained from scratch on the challenge data, with additional post-processing steps, and sometimes using an ensemble of various heavy architectures (ensemble of networks) to reach the highest possible score.…”

One model is all you need: Multi-task learning enables simultaneous histology image segmentation and classification

“…Here, the connective tissue is a broader category that includes endothelial cells, fibroblasts and muscle cells. When creating the dataset, image regions are extracted from the following 6 sources: CRAG (Graham et al, 2019a), GlaS (Sirinukunwattana et al, 2017), CoNSeP (Graham et al, 2019c), PanNuke (Gamper et al, 2020a), DigestPath (Da et al, 2022) and TCGA. Therefore, Lizard is a diverse dataset and models trained on it may likely generalise to unseen examples.…”

Leveraging medical Twitter to build a visual–language foundation model for pathology AI

“…The copyright holder for this preprint (which this version posted April 1, 2023. ; https://doi.org/10.1101/2023.03. 29.534834 doi: bioRxiv preprint training phase, the PLIP model generates two embedding vectors from both the text and image encoders (Figure 1e). These vectors were then optimized to be similar for each of the paired image and text vectors and dissimilar for non-paired images and texts via contrastive learning (Figure 1f, Online Methods).…”

“…In this study, we conducted a systematic evaluation of PLIP's zero-shot capability across four external validation datasets: (i) the Kather colon dataset [24] with 9 different tissue types; (ii) the PanNuke dataset [28] with two tissue types (benign and malignant); (iii) DigestPath dataset [29] with two tissue types (benign and malignant); and (iv) WSSS4LUAD dataset [30] with two tissue types (tumor and normal) (Figure 2b, Extended Figure 2). To evaluate the PLIP model on those datasets, labels were converted to sentences.…”

“…To gain a deeper understanding of the capabilities of the PLIP image encoder, we used four different testing datasets (Kather colon [24], PanNuke [28], DigestPath [29], and WSSS4LUAD [30]) to evaluate the ability of image representation. Image embeddings were first calculated by the PLIP image encoder and then applied dimensionality reduction via UMAP (Figure 3a-d).…”