Ovarian cancer has the second highest incidence and the highest mortality rates among gynecologic malignancies in China. Epithelial ovarian cancer (EOC) accounts for 90-95% of all ovarian cancers. The standard treatment strategy for EOC consists of surgical tumor debulking and administration of intravenous chemotherapy. 

Researchers led by GAO Xin at the Suzhou Institute of Biomedical Engineering and Technology of the Chinese Academy of Sciences applied deep learning to EOC segmentation and evaluated its feasibility. 

"Tumor assessment before surgery is of paramount importance to guide treatment decisions and can ultimately improve patient outcomes," said GAO. Identification and depiction of the tumor area (EOC segmentation) is a prerequisite for tumor assessment, which also facilitates subsequent efficacy evaluation. 

However, EOC segmentation is typically performed slice by slice in clinical practice, which is both time-consuming and labor-intensive. Therefore, a solution to this problem is urgently needed. 

"With the rapid growth of medical image data, the need for fully automated EOC segmentation methods is becoming more and more urgent," said HU Dingdu, first author of the study. 

GAO and his team used a total of 339 magnetic resonance images of EOC patients from eight hospitals in their study. Five evaluation metrics were selected to assess the segmentation performance of different models.

Their results showed that U-Net++ achieved optimal segmentation results in both internal and external test sets demonstrating the superior accuracy and generalization of the model. 

In addition, the team further analyzed the effects of tumor stage and histological type on the segmentation performance of the models. 

However, the segmentation accuracy of the model for advanced-stage tumors and serous tumors is relatively low, "indicating that different tumor stages and histological types have an influence on the segmentation performance of the model," said GAO. 

Visual comparison of segmentation results of different models. (Image by SIBET)