基于深度学习的乳腺癌细胞图像检测算法研究——融合超分辨率增强与注意力机制的YOLO11改进框架

Research on Deep Learning-Based Algorithm for Breast Cancer Cell Image Detection—YOLO11 Enhanced by PFT-SR Super-Resolution, CoordAttention, and Loss Function Optimization

摘要: 乳腺癌有丝分裂细胞计数是病理诊断与预后评估的重要依据，也是衡量肿瘤增殖活性和治疗效果的重要指标。当前人工计数方法受限于主观性强、效率低且重复性差，难以满足临床精准诊断需求。为提升病理图像中小尺度、多形态有丝分裂细胞的检测性能，本文构建了一种融合超分辨率增强与注意力机制的YOLO11改进框架。该方案首先通过PFT-SR模型对病理图像进行超分辨率重建，有效恢复细胞核的细节特征；在此基础上，将CoordAttention注意力机制引入检测网络，增强模型对关键区域的聚焦能力。在GZMH乳腺病理数据集上的实验表明，本文方法在关键指标上达到mAP@0.5：0.597，F1分数：0.608，性能优于主流检测模型，验证了其在有丝分裂细胞检测任务中的有效性与鲁棒性，并展示了方法在临床辅助诊断中的潜在应用价值。

Abstract: Mitotic cell count is a critical parameter in breast cancer for pathological diagnosis, prognosis assessment, and evaluation of tumor proliferative activity. Manual counting, however, is hampered by subjectivity, low throughput, and poor reproducibility, limiting its utility in precision medicine. To address the challenge of detecting small and morphologically diverse mitotic figures in whole-slide images, this study presents an enhanced YOLO11 framework that integrates super-resolution reconstruction and a channel-wise attention mechanism. Our method employs a PFT-SR model to initially restore fine nuclear details through super-resolution, followed by the incorporation of the CoordAttention mechanism to sharpen the network’s focus on critical cellular regions. Evaluated on the GZMH breast histopathology dataset, the proposed approach achieved a mAP@0.5 of 0.597 and an F1-score of 0.608, surpassing the performance of existing mainstream detectors. These results confirm the framework’s robustness and efficacy in mitotic cell detection and underscore its significant potential as a decision-support tool in clinical diagnostics.