Transfer Learning from Tumors to Organs at Risk for Cervical Cancer Image Segmentation

Brachytherapy is one of the most important treatment modalities for cervical cancer, and the delineation of organs at risk (OARs) plays a crucial role in medical image analysis. Accurate segmentation of OARs such as the bladder, rectum, and sigmoid colon helps in precise radiation treatment planning, ensuring accurate delivery of radiation doses to the tumor region while minimizing the impact on surrounding normal tissues and organs. Achieving accurate segmentation is challenging due to the limited availability of sufficient medical images, especially when utilizing deep learning networks. A deep learning algorithm can accurately segment an organ, and to accurately segment tumors and organs, multiple networks are generally trained. We propose a novel organ segmentation network framework that leverages transfer learning, requiring only a small amount of training to use a pre-trained tumor segmentation network for the segmentation of multiple organs. This approach overcomes the problem of overfitting and demonstrates higher segmentation accuracy and spatial consistency. Segmentation results are evaluated using CT image data from 120 cervical cancer patients. Experimental results indicate that the dual cross-domain transfer learning strategy outperforms other mainstream transfer learning strategies, and our segmentation network achieves better performance compared to other mainstream networks.