Mitosis detection in breast cancer using dual-attention noise-free framework

Mitosis counting is an important diagnostic criterion for breast cancer grading and prognosis. This paper proposes a novel mitosis detection method, the Dual-Attention Noise-Free Mitosis Detection Classification Framework (DAN-Mit), aimed at addressing the issues of noise interference and the limitations of multi-scale attention mechanisms in existing detection methods. The framework first utilizes a Rough Classification Filtering (RCF) module to eliminate noise samples, extracting a reliable dataset for training a noise-free model. Next, a dual-attention mechanism is proposed to effectively integrate global and local information through cross-correlation operations, enhancing the model's feature representation and discrimination capabilities. It is noteworthy that the DAN-Mit framework requires only point-level labels to complete the task, offering strong adaptability and efficiency. Experimental results on the MIDOG2021 dataset demonstrate that the DAN-Mit framework significantly improves mitosis detection performance, validating its superiority and effectiveness.