Multi-Layer Pseudo-Supervision for Histopathology Tissue Semantic Segmentation using Patch-level Classification Labels

Abstract: Tissue-level semantic segmentation is a vital step in computational pathology. Fully-supervised models have already achieved outstanding performance with dense pixel-level annotations. However, drawing such labels on the giga-pixel whole slide images is extremely expensive and time-consuming. In this paper, we use only patch-level classification labels to achieve tissue semantic segmentation on histopathology images, finally reducing the annotation efforts. We proposed a two-step model including a classification and a segmentation phases. In the classification phase, we proposed a CAM-based model to generate pseudo masks by patch-level labels. In the segmentation phase, we achieved tissue semantic segmentation by our proposed Multi-Layer Pseudo-Supervision. Several technical novelties have been proposed to reduce the information gap between pixel-level and patch-level annotations. As a part of this paper, we introduced a new weakly-supervised semantic segmentation (WSSS) dataset for lung adenocarcinoma (LUAD-HistoSeg). We conducted several experiments to evaluate our proposed model on two datasets. Our proposed model outperforms two state-of-the-art WSSS approaches. Note that we can achieve comparable quantitative and qualitative results with the fully-supervised model, with only around a 2\% gap for MIoU and FwIoU. By comparing with manual labeling, our model can greatly save the annotation time from hours to minutes. The source code is available at: \url{https://github.com/ChuHan89/WSSS-Tissue}.

• The paper proposes a Multi-Layer Pseudo-Supervision (MLPS) method for histopathology tissue semantic segmentation using only patch-level classification labels, significantly reducing annotation effort.
• Key technical innovations include Progressive Dropout Attention (PDA) for enriched feature extraction and a Classification Gate Mechanism to mitigate false positives in long-tail categories.
• The proposed model achieves performance comparable to fully-supervised methods (2% reduction in MIoU/FwIoU) with a tenfold decrease in labeling time, offering a practical approach for computational pathology.

Multi-Layer Pseudo-Supervision for Histopathology Tissue Semantic Segmentation

The paper by Han et al. addresses the challenge of semantic segmentation in histopathology images using patch-level classification labels, proposing a strategy that aims to significantly reduce the annotation burden typically associated with such tasks. This approach is particularly relevant in computational pathology where dense pixel-level annotations are both costly and time-consuming due to the large size and complexity of whole slide images.

Main Contributions

The authors introduce a two-step model comprising classification and segmentation phases. In the classification phase, a CAM-based model generates pseudo masks guided by patch-level labels. This is followed by a segmentation phase where a novel method called Multi-Layer Pseudo-Supervision (MLPS) is employed to achieve tissue semantic segmentation, effectively bridging the information gap between pixel-level and patch-level annotations. Significant technical innovations are introduced, including:

• Progressive Dropout Attention (PDA): Designed to progressively exclude the most discriminative regions during training, thus prompting neural networks to leverage non-predominant regions for classification, effectively enhancing the richness of feature extraction.
• Classification Gate Mechanism: Implemented to mitigate false-positive errors particularly in categories with fewer samples, thus addressing the long-tail distribution problem prevalent in medical imaging datasets.

In addition, the paper releases a new weakly-supervised dataset, LUAD-HistoSeg, specifically curated for lung adenocarcinoma, to further facilitate research and evaluation in this domain.

Quantitative and Qualitative Evaluation

The paper reports that the proposed model, even when solely based on patch-level annotations, rivals fully-supervised models by achieving comparable results with only a 2% reduction in performance metrics like MIoU and FwIoU. Importantly, it demonstrates a tenfold reduction in labeling time compared to manual annotation, showcasing the marked efficiency gains achievable by this approach.

Figures provided illustrate the model's ability to generate visually accurate segmentation outputs, achieving high concordance with manual annotations while maintaining accuracy across multiple histopathological tissue types.

Potential Applications and Future Directions

The implications of this research are substantial, offering pathways to streamline histopathological analysis with reduced expert human input, thus accelerating clinical workflows. Moreover, the techniques developed could be extended to other domains within medical imaging where annotation costs are prohibitive.

Future research could explore integrating global image context for further reducing false positives and enhancing segmentation accuracy in ambiguous tissue boundaries. Additionally, expanding the framework's applicability to other types of cancers and histopathological challenges may enhance its utility.

Conclusion

Overall, Han et al.'s work demonstrates meaningful advancements in the field of computational pathology by leveraging patch-level classification labels for effective semantic segmentation. This approach is practical, demonstrating significant potential to simplify the workflows required for diagnostic histopathology, thereby forming a robust foundation for future research and applications in medical image analysis.