Deep Slice-Crossed Network With Local Weighted Loss for Brain Metastases Segmentation

Cognitive deficits occur in up to 90% of patients with brain metastases and can be caused by the tumor itself and whole brain radiation therapy. Latest treatment options to reduce cognitive decline require high-precision metastases segmentation. This asks existing metastases segmentation models to be further improved as their limited ability to micro lesion perception. To reduce the segmentation error on brain tissues and improve the accuracy on micro metastases, a deep slice-crossed network (SCNet) and a local weighted loss are proposed, respectively. The SCNet contains a gated feature fusion module where the “gate” is learned individually from the slice-crossed label to roughly localize metastases. Then, in tumor regions, this gate allows intraslice information to pass for metastases enhancement, while in other regions, it allows interslice information to pass for vascular screening. The local weighted loss is proposed to address the inconsistency in size of intraclass objects. Unlike existing weighting approaches, it calculates the weight of each pixel from its neighborhood pixel distribution to enhance the learning of pixels within small tumors, which can be applied to common losses. To evaluate them, two data sets containing 1000 and 368 patients, respectively, were collated. Experimental results demonstrate that the proposed method perform satisfactorily compared with state-of-the-art brain metastasis segmentation models.