Experts' cognition-driven safe noisy labels learning for precise segmentation of residual tumor in breast cancer

Yongquan Yang, Jie Chen, Yani Wei, Mohammad H. Alobaidi, Hong Bu

Abstract

Precise segmentation of residual tumor in breast cancer (PSRTBC) after neoadjuvant chemotherapy is a fundamental key technique in the treatment process of breast cancer. However, achieving PSRTBC is still a challenge, since the breast cancer tissue and tumor cells commonly have complex and varied morphological changes after neoadjuvant chemotherapy, which inevitably increases the difficulty to produce a predictive model that has good generalization with usual supervised learning (SL). To alleviate this situation, in this paper, we propose an experts' cognition-driven safe noisy labels learning (ECDSNLL) approach. In the concept of safe noisy labels learning, which is a typical type of safe weakly supervised learning, ECDSNLL is constructed by integrating the pathology experts' cognition about identifying residual tumor in breast cancer and the artificial intelligence experts' cognition about data modeling with provided data basis. Experimental results show that, compared with usual SL, ECDSNLL can significantly improve the lower bound of a number of UNet variants with 2.42% and 4.1% respectively in recall and fIoU for PSRTBC, while being able to achieve improvements in mean value and upper bound as well.

Introduction

In the treatment of breast cancer, patients often undergo neoadjuvant chemotherapy (NAC) to shrink tumors before surgery. For pathologists and researchers, precisely segmenting the residual tumor in breast cancer (RTBC) after NAC is critical for determining prognosis. However, NAC causes complex morphological changes in both cancer cells and surrounding tissue, creating significant hurdles for traditional AI models. Identifying exact tumor boundaries post-chemo is difficult even for experts, making high-quality datasets rare. Standard supervised learning (SL) models often fail to generalize to these "messy" clinical samples.

Methodology: The ECDSNLL Framework

To address these challenges, the authors propose Experts’ Cognition-Driven Safe Noisy Labels Learning (ECDSNLL). This framework leverages "Safe Noisy Labels Learning" (SNLL) to utilize easily obtained but imperfect data without compromising model accuracy.

Data Preparation: Pathological Expert Cognition

Pathologists provided two distinct datasets based on specific cognitive rules:

  • LAND (Large Amount of Noisy Data): Created using a "loose" rule to exclude non-tumor areas, resulting in high recall but lower precision.
  • SANFD (Small Amount of Noisy-Free Data): Created using a "strict" rule, labeling only certain tumor areas to ensure high precision.

Learning Paradigm: AI Expert Cognition

The framework employs a Two-Stage Cascade Learning Paradigm (TSCLP) to align with these data types:

  1. Stage 1 (Coarse): A Deep Neural Network (DNN) is initially trained on the large, noisy LAND dataset to learn general tumor features and prioritize recall.
  2. Stage 2 (Fine): The pre-trained model is refined using the small, clean SANFD dataset to sharpen precision while maintaining the recall foundation.
Figure 1: The ECDSNLL framework, illustrating the integration of LAND and SANFD datasets into the Two-Stage Cascade Learning Paradigm.

Experimental Results

The ECDSNLL method was evaluated across several UNet variants, including UNet++, UNeXt, and TransUNet. The results demonstrate that ECDSNLL consistently outperforms standard supervised learning (SL) across multiple metrics.

Figure 4: Quantitative results of the ECDSNLL approach across various DNN architectures, showing improvements in recall and fIoU.

Key findings include:

  • Performance Gains: Compared to standard SL, ECDSNLL improved the lower bound of recall by 2.42% and the foreground Intersection over Union (fIoU) by 4.1%.
  • Top Performer: The standard UNet architecture achieved the highest overall performance with a recall of 88.35% and an fIoU of 79.73%.
  • Generalization: The model showed strong qualitative performance on images captured via physical microscopes, indicating robustness beyond digital scanner data.

Conclusion

ECDSNLL offers a practical way to encode human expert logic into AI training pipelines, effectively "embracing the noise" of complex pathological data. While the results are promising for breast cancer, future work will explore the application of this logic to other morphologically complex cancers and aim for product-level reliability in automated clinical diagnostics.

Summary updated

Abstract

Precise segmentation of residual tumor in breast cancer (PSRTBC) after neoadjuvant chemotherapy is a fundamental key technique in the treatment process of breast cancer. However, achieving PSRTBC is still a challenge, since the breast cancer tissue and tumor cells commonly have complex and varied morphological changes after neoadjuvant chemotherapy, which inevitably increases the difficulty to produce a predictive model that has good generalization with usual supervised learning (SL). To alleviate this situation, in this paper, we propose an experts' cognition-driven safe noisy labels learning (ECDSNLL) approach. In the concept of safe noisy labels learning, which is a typical type of safe weakly supervised learning, ECDSNLL is constructed by integrating the pathology experts' cognition about identifying residual tumor in breast cancer and the artificial intelligence experts' cognition about data modeling with provided data basis. Experimental results show that, compared with usual SL, ECDSNLL can significantly improve the lower bound of a number of UNet variants with 2.42% and 4.1% respectively in recall and fIoU for PSRTBC, while being able to achieve improvements in mean value and upper bound as well.
BibTeX
@article{yang2025experts,
  author = {Yongquan Yang and Jie Chen and Yani Wei and Mohammad H. Alobaidi and Hong Bu},
  title = {{Experts' cognition-driven safe noisy labels learning for precise segmentation of residual tumor in breast cancer}},
  journal = {Journal of Data Science and Intelligent Systems},
  year = {2025},
  doi = {10.47852/bonviewjdsis52024868},
}
Figure