基于改进YOLOv5s的乳腺癌有丝分裂病理图像检测

Pathological Image Detection of Breast Cancer Mitosis Based on Improved YOLOv5s

乳腺癌病理图像的有丝分裂结果过程中,由于形态相近的细胞存在干扰,有丝分裂细胞目标小,难以分割标记,从而限制乳腺癌分级诊疗效率和准确性。因此,提出了一种基于改进YOLOv5s的乳腺癌病理图像检测算法。在骨干网络中加入Transformer结构,增强对图像小目标的检测能力。并通过引入ACMix结构,融合图像特征,提高检测性能,强化卷积神经网络对小目标的注意力机制。在检测头部分添加SK-attention,确保捕捉小目标的准确度。结果显示,改进的YOLOv5s的检测性能较改进前传统模型性能更加优秀,检测准确率达97.12%,能较好识别乳腺癌病理图像有丝分裂细胞,进而为后续诊疗提供决策依据。

In the process of mitotic results of pathological images of breast cancer,due to the interference of cells with similar morphology,the target of mitotic cells is small,and it is difficult to divide and mark,thus limiting the efficiency and accuracy of breast cancer grading diagnosis and treatment.Therefore,an improved YOLOv5s based pathological image detection algorithm for breast cancer was proposed.Transformer structure is added to the backbone network to enhance the detection ability of small image targets.By introducing ACMix structure and fusing image features,the detection performance is improved,and the attention mechanism of convolutional neural network on small targets is strengthened.Add SK-attention to the detection header to ensure the accuracy of capturing small targets.The results showed that the detection performance of the improved YOLOv5s was better than that of the traditional model before the improvement,and the detection accuracy was 97.12%,which could better identify mitotic cells in pathological images of breast cancer,and thus provide decision-making basis for subsequent diagnosis and treatment.