| Part oneBackground and objective:Training deep learning(DL)models to automatically recognize diseases in nasopharyngeal MRI is a challenging task,and optimizing the performance of DL models is known to be difficult.This study aims to develop a method of training anatomical partition-based DL model which integrates knowledge of clinical anatomical regions and divisions in otorhinolaryngology to automatically recognize diseases using nasopharyngeal MRI.Methods:A total of 3085 participants with nasopharyngeal MRI(Nasopharyngeal carcinoma:1823,nasopharyngeal lymphoid hyperplasia:212,lymphoma:114,chordoma:113,craniopharyngioma:223,and 600 participants with normal nasopharynx)aged 14 years and older were included.Full images(512*512)of 3085 patients constituted 100%of the dataset,50%and 25%of which were randomly retained as two new datasets for model training.Two semantic segmentation models for the automatic segmentation of perinasopharyngeal area based on U-net and Deeplab v3 were trained and evaluated.Two new series of images(seg112 image[112*112]and seg224 image[224*224])were automatically generated by the segmentation model with better performance.Four DL models for nasopharyngeal diseases classification based on EfficientNet-B0,Legacy SE-ResNet34,MobileNetV3 Large100,and DenseNet121 were trained and evaluated under the nine datasets(full image,seg112 image and seg224 image,each with 100%dataset,50%dataset and 25%dataset).The interpretability of the established DL models for diseases classification was evaluated by the Grad-CAM score.Dice similarity coefficient(Dice)was used to evaluate the performance of the DL models for automatic segmentation of perinasopharyngeal area.The receiver operating characteristic(ROC)curve was used to evaluate the performance of the DL models for nasopharyngeal disease classification.The mean value of the average area under the curve(aAUC)of all classification models for the 100%,50%,and 25%datasets were calculated.Analysis of variance was used to compare the performance of the models built with full image,seg112 image and seg224 image.Independent sample t-test was used to compare Grad-CAM’s scores.Statistical significance was set at P<0.05.Results:The Dice values of U-net and Deeplab v3 for the segmentation of perinasopharyngeal were 0.805±0.021 and 0.897±0.029,respectively.When the 100%dataset was used for training,the performances of the DL models trained with the seg112 images(aAUC 0.949±0.052),seg224 images(aAUC 0.948±0.053),and full images(aAUC 0.935±0.053)were similar(P=0.611).When the 25%dataset was used for training,the mean aAUC of the DL models that were trained with seg112 images(0.823±0.116)and seg224 images(0.765±0.155)were significantly better than that were trained with full images.The interpretability of the DL models that were trained with the seg112 images(0.735±0.097)were significantly better than those trained with the full images(0.245±0.043).Conclusion:The method of training DL model for automatic recognition of diseases in nasopharyngeal MRI based on anatomical partition can potentially improve the performance of the model and optimize its interpretability using a smaller dataset.Part twoBackground and objective:Original images and manually segmented images of tumor regions are typically used to train predictive models for the prognosis of nasopharyngeal carcinoma(NPC).However,clinically adopted TNM staging,which suggests the predictive value of the peritumoral region,has not been evaluated.The purpose of this study was to explore whether incorporating the peritumoral region to train deep neural networks could improve the performance of the models.Methods:A total of 381 NPC patients who were divided into high-and low-risk groups according to progression-free survival were retrospectively included.Deeplab v3 and U-net were transferred to build segmentation models for the automatic segmentation of the tumor and suspicious lymph nodes.Five new datasets were constructed by expanding 5,10,20,40,and 60 pixels outward from the edge of the segmented region.Inception-Resnet-V2,ECA-ResNet50t,EfficientNet-B3,and EfficientNet-B0 were transferred and trained with the original,segmented,and five expanded datasets to establish the classification models.The receiver operating characteristic curve was used to evaluate the performance of each model.Results:The Dice coefficients of Deeplab v3 and U-net were 0.741 and 0.737,respectively.The average areas under the curve(aAUCs)of deep learning models for classification trained with the original and segmented images and with images expanded by 5,10,20,40,and 60 pixels were 0.717±0.043,0.739±0.016,0.760±0.010,0.768±0.018,0.802±0.013,0.782±0.039,and 0.753±0.014,respectively.The models trained with the images expanded by 20 pixels obtained the highest aAUC,whereas the models trained with the original images and Deeplab segmented images performed the worst.Conclusion:The peritumoral region on the magnetic resonance imaging image of NPC contains information related to prognosis,and the incorporation of this region could improve the performance of deep learning models for prognosis prediction. |
PDF Full Text Request