Malignant Mesothelioma Subtyping of Tissue Images via Sampling Driven Multiple Instance Prediction

MesoGraph: Automatic profiling of mesothelioma subtypes from histological images

“…As can be seen from these results, the proposed approach offers high predictive quality over both cross-validation and independent testing in comparison to other existing approaches. In the table, PINS refers to the positive instance sampling patch-based MIL approach as detailed in Eastwood et al., 13 whereas CLAM is the clustering-constrained attention MIL method described in Lu et al., 14 a deep-learning-based weakly supervised method that uses attention in combination with clustering-based constraints to identify the most predictive areas of the image. Max-MIL and naive-MIL are simple patch-based baseline MIL methods detailed further in the STAR Methods Model performance and evaluation.…”

“…The cores to be used for training are split 75%–25% into train and validation sets, respectively. We compared our model with CLAM, 14 and PINS, 13 two patch-based methods which attempt to focus training in an adaptive way on the most important instances. We additionally compared with two simple MIL approaches, max-MIL and naive-MIL.…”