Background: Medical image segmentation is an important task in the diagnosis and treatment of cancers. The low contrast and highly flexible anatomical structure make it challenging to accurately segment the organs or lesions.
Purpose: To improve the segmentation accuracy of the organs or lesions in magnetic resonance (MR) images, which can be useful in clinical diagnosis and treatment of cancers.
Methods: First, a selective feature interaction (SFI) module is designed to selectively extract the similar features of the sequence images based on the similarity interaction. Second, a multi-scale guided feature reconstruction (MGFR) module is designed to reconstruct low-level semantic features and focus on small targets and the edges of the pancreas. Third, to reduce manual annotation of large amounts of data, a semi-supervised training method is also proposed. Uncertainty estimation is used to further improve the segmentation accuracy.
Results: Three hundred ninety-five 3D MR images from 395 patients with pancreatic cancer, 259 3D MR images from 259 patients with brain tumors, and four-fold cross-validation strategy are used to evaluate the proposed method. Compared to state-of-the-art deep learning segmentation networks, the proposed method can achieve better segmentation of pancreas or tumors in MR images.
Conclusions: SFI-Net can fuse dual sequence MR images for abnormal pancreas or tumor segmentation. The proposed semi-supervised strategy can further improve the performance of SFI-Net.
Keywords: dual sequences; feature interactive fusion; semi-supervision.
© 2022 American Association of Physicists in Medicine.