Research On Generalization Performance Of Brain Tumor Segmentation Supervised Learning Model

Semantic segmentation converts original biomedical image data into mean-ingful and spatially structured information,so it plays a vital role in medical im-age analysis and research.Although the segmentation methods based on deep learning have achieved success recently,they are restricted by the relatively small number of samples in the medical imaging data set,which leads to the unsatisfactory generalization of their models.In particular,the location,size,and shape of brain gliomas are highly specific,which makes it difficult to manu-ally label data,which can easily lead to misjudgments and divergence,resulting in label noise,and poses a higher challenge to generalization.In response to this problem,from the perspective of data enhancement,this paper proposes a data enhancement method tumor-mixup,which is more stable than traditional data enhancement methods,and increases the Dice value of the ET subregion from 0.723 to 0.768.From the perspective of uncertainty,the segmentation model is calibrated through the auxiliary network Calibrate block,so that the model avoids generating prediction results with too high confidence,and reduces the ECE of the ET subregion from 0.28 to 0.243.From the perspective of label noise,a label denoising method called Kickflip is proposed,which enables the network to generate predictions and denoise labels at the same time.When the Calibrate block and Kickflip methods are combined,the ECE can finally be further reduced to 0.217.