Study On The Method Of Automatically Delineating The Target Area Of Nasopharyngeal Carcinoma In Radiotherapy Planning

Nasopharyngeal carcinoma is a common malignant tumor.Radiotherapy is the main treatment for nasopharyngeal carcinoma,and accurate delineation of the primary tumor target area is a prerequisite for radiotherapy.However,manual lesion delineation is cumbersome and error-prone.Therefore,achieving automatic delineation of the primary nasopharyngeal carcinoma target area from computed tomography(CT)images in radiotherapy planning has important clinical value.However,there are still several issues with achieving accurate automated segmentation of the target area.First,the contrast between the primary nasopharyngeal carcinoma and surrounding soft tissues in CT images is low and not easy to recognize.Second,the boundary is fuzzy and occupies a very small proportion in CT images,leading to extreme imbalance between background and segmentation targets.Third,the anisotropy of nasopharyngeal carcinoma CT images causes significant shape changes in the primary lesion target area in adjacent slices.Finally,due to the scarcity of labeled nasopharyngeal carcinoma images,overfitting is prone to occur when the training dataset is insufficient.In response to these issues,the following research is conducted:(1)This article addresses the issues of low contrast,fuzzy boundary,small target area,and anisotropy of nasopharyngeal carcinoma(NPC)in CT images.Using the 2.5D U-Net as the main architecture,two fully supervised semantic segmentation algorithm models were developed: the Modified Channel and Spatial Attention U-Network(MCSA-UNet)and the Fused Attention U-Network(FA-UNet).These models incorporate comprehensive attention mechanisms such as channel,spatial,and self-attention,to solve the anisotropy problem of NPC CT image segmentation.The MCSA-UNet model introduces the MCSA attention module in the encoding and decoding layers to address the small target segmentation problem.A 3D crossattention module based on self-attention is also introduced in the bottleneck layer to solve the problem of low contrast between the target area and the surrounding soft tissue.To address the problem of fuzzy edges,an attention gate is introduced in the decoder to further compensate and optimize relevant edge feature information.On both public and clinical datasets,compared to the classical 3D U-Net,the Dice coefficient of MCSA-UNet is improved by 7.93% and9.78%,and the Hausdorff distance is improved by 34.86 mm and 50.30 mm,respectively.The FA-UNet model further improves the MCSA attention module by incorporating branch structures,frequency domain methods,and dilated convolutions in the encoding and decoding layers.In the bottleneck layer,a 3D interleaved sparse attention module combining group convolution and improved self-attention is introduced.Similar to the MCSA-UNet,the FAUNet also solves the problem of fuzzy edges by introducing an attention gate in the decoder.On both public and clinical datasets,compared to the 3D U-Net,the Dice coefficient of FAUNet is improved by 9.84% and 11.01%,and the Hausdorff distance is improved by 35.89 mm and 52.58 mm,respectively.(2)This passage describes a method for addressing the problem of scarce labeled nasopharyngeal carcinoma(NPC)images using a semi-supervised learning strategy,called the Non-relevance Consistency Model(NCM).The framework includes both student and teacher models,with an objective optimization function that combines supervised and consistency losses.A novel consistency scheme is designed to guide the student model towards more reliable information using an uncertainty-guided strategy.The proposed NCM model is compared to other state-of-the-art semi-supervised segmentation methods,and under the condition of 50% labeled data,the NCM method achieves higher Dice coefficient(70.08%)and Hausdorff distance(HD)(4.86mm)on a public dataset.On a clinical dataset,the NCM method also outperforms other comparison methods with higher Dice coefficient(74.29%),recall(77.46%),and HD(4.08mm).