Data Uncertainty Learning in Breast Cancer Recognition
Ku Zhao, Tao Luo, Kehan Chen, Libo Zhang · 2023
Breast cancer has emerged as the most prevalent malignancy, occupying the top position in both incidence and mortality among cancers in women. Distinguishing between benign and malignant breast tumors plays a crucial role in achieving early detection and treatment of breast cancer. As breast biopsy is a critical diagnostic evidence for confirming the malignant tumors of the breast, using its pathological images to diagnose the benign or malignant nature of breast tumors by deep learning models has become an important approach in breast cancer recognition. The noise in the digitization process of biopsy samples' pathological images will lead to decision uncertainty. Nevertheless, most existing deterministic models fail to accurately represent the uncertainty in the data, leading to incorrect breast cancer diagnoses. Misdiagnosis comes at a high cost, as it can result in patients missing the optimal treatment window or receiving unnecessary treatments, etc. To solve this problem, we introduce data uncertainty learning, an effective method for representing noise, into breast cancer recognition for the first time. Then, we propose a data uncertainty learning model for breast cancer recognition task, which learns the mean to recognize breast cancer and variance to represent uncertainty of samples. Specifically, we design a variance regularizer to guide the model adjusting optimization weights of the samples based on the uncertainty and enable the model to focus more on high-quality samples while suppressing the influence of noise, which enhances the performance of breast cancer recognition. Finally, insightful evaluation demonstrates the effectiveness of our model.