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 label learning (ECDSNLL) approach. In the concept of safe noisy label learning, which is a typical type of safe weakly SL, 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

Following neoadjuvant chemotherapy (NAC), assessing residual tumor in breast cancer is critical for determining patient prognosis and subsequent treatment paths. However, NAC induces complex morphological changes—such as tumor shrinkage and tissue scarring—that make precise segmentation (PSRTBC) extremely difficult for standard artificial intelligence models. Traditional supervised learning (SL) requires high-quality, pixel-perfect labels, which are labor-intensive to produce at scale. To address this, researchers have developed the Experts’ Cognition-driven Safe Noisy Label Learning (ECDSNLL) framework, which strategically combines expert pathology intuition with advanced AI training paradigms.

Methodology

The ECDSNLL approach utilizes a "Safe" weakly supervised learning paradigm, ensuring that the inclusion of noisy data does not degrade model performance compared to using clean data alone. The framework is built upon two pillars of expert cognition:

  1. Pathology Expert Cognition (Data Preparation): Pathologists provided two distinct datasets. The LAND (Large Amount of Noisy Data) set consists of coarse labels prioritizing high recall to ensure all potential tumor regions are captured. The SANFD (Small Amount of Noisy-Free Data) set contains a smaller volume of high-precision, pixel-perfect labels.
  2. AI Expert Cognition (Two-Stage Cascade Learning Paradigm): As illustrated in Figure 3, the model undergoes a two-stage training process. In Stage 1 (S1), the model learns broad morphological features from the LAND dataset. In Stage 2 (S2), the model is fine-tuned using the SANFD dataset to refine boundaries and improve precision.
The ECDSNLL framework integrates pathology and AI expert cognition through a two-stage cascade learning paradigm.

Results and Discussion

The framework was tested across several UNet variants (e.g., UNet++, UNet3+, AttUNet). Compared to standard supervised learning, ECDSNLL significantly improved the "worst-case scenario" performance, raising the lower bound of recall by 2.42% and foreground Intersection over Union (fIoU) by 4.1%. The top-performing variant, UNet(ECDSNLL), achieved a recall of 90.89% and an fIoU of 81.88%.

Precision-Recall curves comparing the proposed ECDSNLL approach against standard supervised learning (SL).

As shown in the Precision-Recall curves (Figure 5), the ECDSNLL approach consistently maintains higher precision across various recall levels than the SL baseline. This demonstrates superior generalization, even when tested on "out-of-distribution" data such as images captured directly through microscope eyepieces rather than digital scanners.

Conclusion

ECDSNLL offers a robust solution for medical image segmentation tasks where high-quality labels are scarce but coarse labels are abundant. By leveraging the strengths of both noisy and clean data through a structured two-stage process, the framework achieves higher reliability in identifying residual breast cancer tumors. Future research will focus on expanding the dataset volume and applying this framework to other cancer types affected by chemotherapy-induced tissue changes.

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 label learning (ECDSNLL) approach. In the concept of safe noisy label learning, which is a typical type of safe weakly SL, 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. Received: 21 November 2024 | Revised: 10 January 2025 | Accepted: 23 January 2025 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement Data available on request from the corresponding author upon reasonable request. Author Contribution Statement Yongquan Yang: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing – original draft, Visualization, Supervision, Project administration. Jie Chen: Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – review & editing. Yani Wei: Validation, Formal analysis, Investigation, Resources, Data curation, Writing – review & editing. Mohammad Alobaidi: Software, Validation, Formal analysis, Investigation, Writing – review & editing. Hong Bu: Validation, Formal analysis, Investigation, Resources, Writing – review & editing, Supervision, Project administration, Funding acquisition.
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