Fusion attention-based nasopharyngeal carcinoma segmentation model in predicting the clinical outcome of cervical...

Deep learning methods have made great progress in the automatic segmentation of nasopharyngeal carcinoma, but challenges remain.

Computer-aided automatic segmentation of nasopharyngeal cancer primary area is of great significance for automatic outlining of nasopharyngeal cancer target areas and accurate prediction of responsiveness and prognosis of metastatic lymph nodes in the neck after radiotherapy. In this paper, we use deep learning methods to construct an automatic segmentation network for gross target volume of nasopharynx, combine clinical factors and radiomics features to establish a radiomics nomogram model, which will then predict the final outcome of metastatic lymph nodes that have not achieved complete remission after radical radiotherapy.

Clinical and IMRT radiotherapy plan CT data were retrospectively collected from 69 patients who received intensity-modulated radiation therapy between July 2014 and December 2016. These patients exhibited residual metastatic lymph node lesions without residual primary lesions on the first follow-up MRI and had continuous follow-up records. The median follow-up was 53 months (IQR 39.75-62.37), with 30 patients eventually regressing and 39 patients persisting or progressing. The ct images of 69 radiotherapy plans were randomly divided into training and test sets according to 8:2, and a fusion attention-based model was trained for automatic nasopharyngeal carcinoma segmentation. Based on the unet framework, a fusion attention model was proposed, and a 2·5 d convolutional neural network was used to deal with the anisotropy. An improved channel and spatial attention module is fused in the codec 4 layer to enable the network to focus on small targets. 2d interlaced sparse self-attention module is extended to 3d to better extract the feature information of the tumor target area and solve the problem of low contrast between the target area and the surrounding soft tissues, thus optimizing the overall segmentation effect. The performance of the segmentation model was evaluated using the mean dice coefficient, relative volume error (RVE), average symmetric surface distance (ASSD) and hausdorff distance (HD), using the target area of the primary lesion of nasopharyngeal carcinoma manually outlined by a senior radiation therapy specialist as the gold standard. Radiomics features were extracted using the pyradiomics package, and the classification performance of the radiomics model was assessed by the area under the curve of the receiver operating curve (ROC).

The average dice coefficient, RVE, ASSD and HD of our model for nasopharyngeal carcinoma were 75.05%, 14.63%, 2.224 mm, and 8.75 mm, respectively, which were 11.01%, 26.34%, 3.101 mm, and 52.58 mm better than the baseline 3dunet model. The radiomic features were an effective predictor of tumor outcome in nasopharyngeal carcinoma, with the highest area under the receiver operating characteristic curve (AUC) of 0.892 for the radiomic nomogram in the training set and 0.825 for the radiomic model in the test set.

The fused attention-based segmentation network for nasopharyngeal carcinoma can effectively and reliably segment the region of the primary nasopharyngeal carcinoma, and the radiomic nomogram can effectively predict the response after treatment.