Abstract:
Abstract： High grade serous ovarian cancer is a malignant tumor disease， and preoperative recurrence prediction can help clinical doctors provide personalized treatment plans for patients and reduce the mortality rate. Due to the less and difficult-to-obtain medical data of this disease， its deep learning model is difficult to obtain sufficient training， and the accuracy of recurrence prediction needs to be improved. To address this issue， this article proposes an improved low-parameter residual network TGE-ResNet34， which uses ResNet34 as the backbone network and replaces traditional convolution modules with Ghost convolutions to extract lesion area features and reduce the model’s parameter volume. The ECA （Efficient Channel Attention） attention mechanism is incorporated between two Ghost convolutions to suppress interference from useless feature extraction. Finally， the model is evaluated through a five-fold cross-validation to avoid the randomness of data partitioning. The experimental results show that the accuracy of the improved TGE-ResNet34 network is 96.01%， which is 4.52 percentage points higher than the original baseline network’s accuracy and reduces the parameter volume by 15.98 M.

Key words: Key words： high grade serous ovarian cancer, residual network, Ghost convolution, attention