A Research On Weakly Supervised Automatic Segmentation Of Brain Tumor Images

Automatic segmentation of brain tumors from 3D medical images is critical for tumor diagnosis and treatment.With the advancement of deep learning,Convolutional Neural Networks(CNNs)have achieved significant success in 3D medical image segmentation.However,training CNN requires a large amount of images with corresponding pixel-level annotations.Medical images often have privacy protection and ethical constraints,making data acquisition challenging,and annotation requires expertise from medical professionals,resulting in high time and labor costs.Weakly supervised medical image segmentation,which utilizes sparse weak annotations to train CNNs,can address the challenge of obtaining dense annotations.However,weakly supervised medical image segmentation remains challenging,as it suffers from limited supervision signals provided by weak annotations and prediction results of the model often contain errors(noise).To address these issues,this study proposes a novel weak annotation method and a two-stage weakly supervised brain tumor segmentation framework called PA-Seg(Point Annotations Segmentation).The main content and innovative points of this article are as follows:(1)The paper proposes a novel point annotation method,which only requires 7points to annotate the tumor region in 3D medical images,and the accuracy requirement for the annotation points is not high,making it more user-friendly for annotators without medical knowledge.Based on this annotation method,PA-Seg is used to train powerful segmentation models,effectively addressing the issue of scarce data annotation in the field of medical imaging.(2)The first stage of PA-Seg aims to train an initial segmentation model from point annotations.In this study,geodesic distance transform is employed to expand the seed points of the annotations,providing more supervision signals.To further handle the unlabeled image regions during training,two context regularization terms are proposed:multi-view conditional random field loss and variance minimization loss.The former encourages pixels with similar features to have consistent labels,while the latter minimizes the inter-class pixel value variance in the segmentation results.(3)The goal of the second stage of PA-Seg is to reduce errors in model predictions.In this study,the model trained in the first stage is used to predict segmentation as pseudolabels.To overcome the noise in pseudo-labels,an auxiliary network is introduced alongside the main network,and a Self and Cross-Monitoring(SCM)approach is proposed for training both networks.SCM consists of two parts: self-training,where the main and auxiliary networks train themselves using their respective generated hard labels,and cross knowledge distillation,where the main and auxiliary networks learn from the soft labels generated by each other,reducing noise in the predicted results.Experiments on Vestibular Schwannoma(VS)segmentation and Brain Tumor Segmentation(BraTS)datasets demonstrate that the model trained in the first stage significantly outperforms existing state-of-the-art weakly supervised segmentation methods.The Dice coefficients on the VS and BraTS datasets are 0.836 and 0.842,respectively.After additional training using the second stage SCM,the Dice coefficients improve to 0.852 and 0.856,respectively,representing an improvement of 15.5% and 13.8% compared to the baseline methods.The performance of the model on the BraTS dataset approaches that of a fully supervised model.