CASC-AI: Consensus-aware Self-corrective Learning for Noise Cell Segmentation

TL;DR

The paper tackles multi-class cell segmentation in gigapixel WSIs when training data come from lay annotators and contain noise. It introduces CASC-AI, a Consensus-aware Self-Corrective Learning framework that uses a Consensus Matrix to identify CP, CN, DM, and DH regions and employs both consensus-guided supervision and a contrastive noise separation objective to correct noisy labels. Key contributions include the Consensus Matrix design, a three-component learning paradigm, and reasoning-guided noise generation, all leading to improved segmentation performance under noisy supervision. The approach demonstrates robust FP/FN correction and improved generalization on real lay-annotated data as well as simulated noise datasets, offering a scalable path for training cell-segmentation models with weak supervision in digital pathology.

Abstract

Multi-class cell segmentation in high-resolution gigapixel whole slide images (WSIs) is crucial for various clinical applications. However, training such models typically requires labor-intensive, pixel-wise annotations by domain experts. Recent efforts have democratized this process by involving lay annotators without medical expertise. However, conventional non-corrective approaches struggle to handle annotation noise adaptively because they lack mechanisms to mitigate false positives (FP) and false negatives (FN) at both the image-feature and pixel levels. In this paper, we propose a consensus-aware self-corrective AI agent that leverages the Consensus Matrix to guide its learning process. The Consensus Matrix defines regions where both the AI and annotators agree on cell and non-cell annotations, which are prioritized with stronger supervision. Conversely, areas of disagreement are adaptively weighted based on their feature similarity to high-confidence consensus regions, with more similar regions receiving greater attention. Additionally, contrastive learning is employed to separate features of noisy regions from those of reliable consensus regions by maximizing their dissimilarity. This paradigm enables the model to iteratively refine noisy labels, enhancing its robustness. Validated on one real-world lay-annotated cell dataset and two reasoning-guided simulated noisy datasets, our method demonstrates improved segmentation performance, effectively correcting FP and FN errors and showcasing its potential for training robust models on noisy datasets. The official implementation and cell annotations are publicly available at https://github.com/ddrrnn123/CASC-AI.

Figures (4)

• Figure 1: Consensus-aware self-corrective learning. We propose a Consensus-Aware Self-Corrective Learning for robust cell segmentation with noisy training data. The model leverages the CM to guide learning, prioritizing CP and CN regions with stronger supervision, while adaptively weighting DM and DH regions based on their similarity to reliable CP regions by contrastive learning.
• Figure 2: Overview of the Consensus-Aware Supervision Framework. The architecture integrates AI-derived confidence maps ($c$) and lay annotations ($y_l$) to identify consensus-positive (CP), consensus-negative (CN), and disagreement regions (DM, DH). This framework emphasizes robust training by focusing on regions of consensus and leveraging disagreement as informative cues for improved cell segmentation accuracy.
• Figure 3: Illustration of the Noisy Dataset. The figure depicts a real lay annotation dataset and two reasonable noise generation pipelines used to create FP and FN datasets with plausible noise. These processes are applied to evaluate the proposed method under challenging scenarios.
• Figure 4: Qualitative Results. The figure presents qualitative results on real dataset obtained using the supervised method and the proposed CASC-AI method. The results demonstrate that the proposed self-corrective learning approach enhances segmentation performance on noisy labels by reducing false positives and false negatives.