基于背景激活抑制的弱监督脑肿瘤图像分割算法OA

Magnetic resonance imaging(MRI)is a commonly used medical technique for brain tumor detection.The accurate brain tumor segmentation is crucial for assessing patient conditions and formulating treatment plans.This paper proposes a background activation suppression mechanism to deal with the excessive background activation caused by unclear subject matter in brain tumor images.Firstly,a classifier is trained to extract the features.Secondly,a background activation suppression module is designed and embedded into the backbone network of the classifier.This module effectively adjusts attention from background regions to foreground regions,and a preliminary segmentation mask is obtained.Finally,a boundary refinement module is introduced.The boundaries of segmentation masks are further refined by calculating affinities from both RGB image features and spatial location information,so that a more precise segmentation is obtained.The proposed method is evaluated on four modalities of the BraTS 2021 dataset and MSD Brain dataset.On the T2-FLAIR modality of the BraTS 2021 dataset,its Dice coefficient and intersection over union(IoU)reach 0.845 and 0.788,respectively.These results are 6.0%and 7.2%higher than those of the Cfd-CAM,respectively,outperforming the existing algorithms.The algorithm presented in this paper improves the accuracy of brain tumor segmentation.It provides valuable ideas for other weakly supervised semantic segmentation(WSSS)algorithms for medical images.