One of the key challenges in medical image segmentation is the weak boundaries of organs and structures. To address this challenge, we propose a novel demarcation line/ridge detection model using Mask R-CNN, which is able to accurately detect the boundaries of organs and structures with weak edges. This model is applied to the detection of Retinopathy of Prematurity (ROP) stage 2, a critical step in the diagnosis and treatment of ROP.

Another challenge in medical image segmentation is the variability of medical images due to factors such as patient anatomy, image quality, and modality. We propose an efficient liver segmentation model that combines holistically nested edge detection (HED) and Mask R-CNN to address this challenge. This model is able to accurately segment the liver from both CT and MRI images, even in the presence of noise and artifacts. We also propose an efficient spine vertebrae segmentation model using Mask R-CNN, combined with a complete intersection over union (CIoU) and active contour loss. This model is able to accurately segment the spine vertebrae from CT images and can be generalized to MRI images.
Finally, we propose an adaptive instance normalization style transfer method using the dense extreme inception network and convolution block attention module (CBAM) to achieve the best vessel segmentation performance. This model is trained on natural images but can be applied to medical images without the need for a large dataset or labeled ground truth. The proposed models can potentially improve the accuracy and efficiency of medical image segmentation and facilitate the development of new clinical applications.