Researchers present MitPro, an AI tool that directs pathologists to regions of high predicted mitotic activity and highlights mitotic figures while keeping humans in control. In a 385-slide paired reader study across 3 centres and 7 tumour types, AI assistance raised inter-rater agreement (ICC 0.589 to 0.949) and nearly halved assessment time. Results support assistive use for tumour grading in routine practice.
A new arXiv paper presents MiTHras, a pretraining framework combining pseudo-label-guided image- and token-level contrastive learning with masked reconstruction for mitotic figure analysis. The authors also release TCGA-MF-Pseudo, a 1.8-million-image corpus built from 14 TCGA cohorts. MiTHras achieves top mean F1 across MF classification and subtype benchmarks, improving robustness of automated tumor grading tools.
A new arXiv paper presents MiTHras, a pretraining framework combining pseudo-label-guided image- and token-level contrastive learning with masked reconstruction for mitotic figure analysis. The authors also release TCGA-MF-Pseudo, a 1.8-million-image corpus built from 14 TCGA cohorts. MiTHras achieves top mean F1 across MF classification and subtype benchmarks, improving robustness of automated tumor grading tools.
A new arXiv paper argues that quality-control benchmarks for whole-slide image analysis produce hard-to-interpret results because of four structural flaws: few independent slides, annotations concentrated in a minority of slides, pooled ratio metrics lacking closed-form standard errors, and a single inherited train/test partition. The authors propose a reliability protocol that quantifies variability from test-set sampling, training stochasticity, partition composition, and uncertainty. It matters to practitioners building artifact-detection pipelines who need reproducible benchmark comparisons.
A new arXiv paper argues that quality-control benchmarks for whole-slide image analysis produce hard-to-interpret results because of four structural flaws: few independent slides, annotations concentrated in a minority of slides, pooled ratio metrics lacking closed-form standard errors, and a single inherited train/test partition. The authors propose a reliability protocol that quantifies variability from test-set sampling, training stochasticity, partition composition, and uncertainty. It matters to practitioners building artifact-detection pipelines who need reproducible benchmark comparisons.
Researchers propose Loki-OT, a method that propagates region-level tissue reasoning to individual cell predictions via Unbalanced Optimal Transport, using MLLM-derived density priors as soft guidance for ambiguous cell reassignment. This addresses cell mimicry, where different cell types appear morphologically similar, by leveraging surrounding tissue context that current vision models lack.
Researchers propose Loki-OT, a method that propagates region-level tissue reasoning to individual cell predictions via Unbalanced Optimal Transport, using MLLM-derived density priors as soft guidance for ambiguous cell reassignment. This addresses cell mimicry, where different cell types appear morphologically similar, by leveraging surrounding tissue context that current vision models lack.
An arXiv paper evaluates whether pathology foundation models encode acquisition-center information that leaks into whole-slide image predictions. Using controlled training with increasing class-center correlations measured by Cramér's V, the authors benchmark six PFMs across four datasets and two MIL aggregators. Results show center-related information does propagate to slide-level predictions, with robustness varying by model and aggregation strategy, and ComBat harmonization failing to consistently help.
An arXiv paper evaluates whether pathology foundation models encode acquisition-center information that leaks into whole-slide image predictions. Using controlled training with increasing class-center correlations measured by Cramér's V, the authors benchmark six PFMs across four datasets and two MIL aggregators. Results show center-related information does propagate to slide-level predictions, with robustness varying by model and aggregation strategy, and ComBat harmonization failing to consistently help.
A new arXiv paper, StainPresetNet, presents a stain preset network for fast multi-to-multi stain normalization in computational pathology. It targets limitations of traditional pixel-wise mapping methods and prior deep-learning approaches, which are either inaccurate in color mapping extraction or computationally inefficient. Stain normalization matters to practitioners because it reduces staining-protocol and imaging-condition variability, improving computer-aided diagnostic performance.
A new arXiv paper, StainPresetNet, presents a stain preset network for fast multi-to-multi stain normalization in computational pathology. It targets limitations of traditional pixel-wise mapping methods and prior deep-learning approaches, which are either inaccurate in color mapping extraction or computationally inefficient. Stain normalization matters to practitioners because it reduces staining-protocol and imaging-condition variability, improving computer-aided diagnostic performance.