Key point detection of breast pose estimation based on DeepLabCut

Evaluating the aesthetic effect on patient breasts after oncoplastic breast surgery using computer technology can effectively help surgeons objectively analyse and judge the surgical effect. Many software programs are available at present to accurately evaluate the effect of surgery through postoperative photos. However, existing software generally requires users to mark key points in photos, which greatly increases the workload. This study uses deep neural network techniques for breast key point detection to address the shortcomings of existing software, reduce workload, and improve efficiency. We design a DeepLabCut network from three aspects: a pretrained residual network, transposed convolution, and a loss function. We trained the model using photos of more than 300 breast surgery patients from Tianjin Medical University Cancer Institute & Hospital. The results showed that when the threshold was set to 9 pixels, the accuracy of DeepLabCut in detecting the key points of the front, left, and right sides of the breast was 88.11%, 91.02%, and 92.55%, respectively. Furthermore, we compared our model with a popular convolutional pose machine (CPM) and stacked hourglass and found that our model has higher accuracy. Our work provides the foundation for developing new software for evaluating the effectiveness of breast surgery.