SeqSeg: A sequential method to achieve nasopharyngeal carcinoma segmentation free from background dominance

COVID-19 diagnosis via chest X-ray image classification based on multiscale class residual attention

“…Attention mechanisms are extensively adopted in computer vision community, such as classification [ [67] , [68] , [69] , [70] , [71] , [72] ], detection [ [73] , [74] , [75] , [76] ] and segmentation [ 28 , [77] , [78] , [79] ]. On the basis of deep learning, CXR image classification can discriminates different pathologies by feature learning.…”

Add-on individualizing prediction of nasopharyngeal carcinoma using deep-learning based on MRI: A multicentre, validation study

“…Next, the performance of SeqSeg was evaluated on the large nasopharyngeal carcinoma dataset containing 1101 patients. The experimental results demonstrated that the proposed SeqSeg not only outperforms several state-of-the-art methods but also achieves better performance in multi-device and multi-centre datasets ( Tao et al., 2022 ). The application of NPCnet and Seqseg to segment primary tumor and/or lymph nodes as markers for nasopharyngeal carcinoma patient staging is beneficial in prediction and radiotherapy planning.…”

Deep learning in MRI‐guided radiation therapy: A systematic review

“…Notable efforts include the two-stage multi-channel Seqseg architecture for NPC segmentation. 71 Seqseg uses reinforcement learning to refine the position of the bounding box, implements residual blocks, recurrent channel and region-wise attention, and a custom loss function that emphasizes segmentation of the edges of the tumor. Outierial et al 72 improves the dice score by 0.10 with a two-stage approach compared to single-state 3D U-Net for oropharyngeal cancer segmentation.…”

“…Focal loss is the cross‐entropy loss modified for increased sensitivity 42 and Tversky loss does the same for the dice loss 43 . In addition, borders of the contours are the most important part of the segmentation, so boundary loss functions seek to improve model performance by placing increased emphasis on regions near the contour edge 44,45 . Another approach to solve the problem, albeit at the expense of long‐range context, is with two stage networks.…”