A Research On Focus Segmentation And Prognosis Of Patients With Nasopharyngeal Carcinoma From Mr Images Based On Deep Learning

Nasopharyngeal carcinoma(NPC)is the most common malignant tumor in the human nasopharynx.The detection and accurate localization of tumors and lymph nodes in medical images are important basis for diagnosis,treatment planning and prognosis.A total of 647 patients with nasopharyngeal carcinoma subjects were selected from Sun Yat-sen University Cancer Center,the clinical information and three sequences of head and neck MR images,T1 W,T2W and T1 C,were collected.The tumors and lymph nodes were delineated by experienced doctors as ground truth.All subjects were divided into training sets and independent test sets with the ratio of 3:1.According to the characteristics of multi-sequence MR images,the multi-sequence 2D-ResUNet and the multi-sequence and multi-dimensional fusion(MSMDF)models were constructed to segment tumor regions and lymph nodes.Based on the segmentation results,image features were extracted and then combined with clinical indicators such as TN staging,cranial nerve invasion,EBV-DNA copy number,pre-treatment EBV content,VCAIgA,EAIgA to construct a classification prediction model,so as to evaluate the postoperative metastasis and recurrence.The experimental results show that the Dice values of the multi-sequence 2D-ResUNet for tumors and lymph nodes segmentation are 0.786 and 0.808,respectively,and the Hausdorff distance(HD)is 6.09 mm and 5.72 mm,the percentage of area difference(PAD)was 19.1% and 17.7%,respectively.Whereas the Dice values of the MSMDF model for tumors and lymph nodes segmentation were 0.801 and 0.830,HD was 5.97 mm and 5.43 mm,and PAD was 18.0% and 15.6%,respectively.For the metastasis evaluation,the AUC values of models using clinical indicators,image features,and their combination were 0.733,0.791,and 0.840,respectively,and the F1 values were 0.458,0.622,and 0.727,respectively.For the recurrence evaluation,the AUC values of the models using clinical indicators,image features,and their combination were 0.727,0.802,and 0.849,respectively,and the F1 values were 0.432,0.571,and 0.686,respectively.This study explored the deep learning models to fulfill the automatic segmentation of NPC tumors and lymph nodes,and the evaluation of NPC tumor metastasis and recurrence.Experimental results demonstrated that the proposed models could detected tumors and lymph nodes accurately,and the combination of segmentation results and clinical indicators could be applied for further postoperative metastasis and recurrence evaluation.The proposed methods and results have great potentials in the diagnosis,treatment and prognosis of NPC patients.