3D-IncNet: Head and Neck (H&N) Primary Tumors Segmentation and Survival Prediction

Multi-modal tumor segmentation methods based on deep learning: a narrative review

“…The networks are categorized into input-level fusion and feature-level fusion, and the table includes information on the network structure, preprocessing method, dataset used, and common segmentation result indicators. Qayyum et al validated the HECKTOR21 and HECKTOR22 datasets in their work ( 60 , 86 ). They fused PET and CT 2-modality images at the input-level and then performed segmentation using a U-shaped segmentation network.…”

“…In 2014, Google proposed a network structure called GoogleNet ( 58 ), which included the Inception module. In multi-modal tumor segmentation tasks ( 40 , 59 , 60 ), the Inception module has also been used as a component of the structural framework. For example, Qayyum et al ( 60 ) used the Inception structure as the encoding part in the HNSCC segmentation task.…”

“…In multi-modal tumor segmentation tasks ( 40 , 59 , 60 ), the Inception module has also been used as a component of the structural framework. For example, Qayyum et al ( 60 ) used the Inception structure as the encoding part in the HNSCC segmentation task. They extracted features of different levels to avoid gradient disappearance and reduce computational costs, resulting in improved segmentation performance.…”

“…Compared with traditional weight transfer methods, such as transferring feature maps to the next layer, SENet establishes interdependence between channels and adaptively corrects feature strengths among channels through a global loss function. Some methods ( 25 , 42 , 43 , 60 , 86 - 88 ) integrate SENet into their own network to improve feature extraction abilities. For example, Qayyum et al ( 60 , 86 ) embedded SENet into the Inception coding part to obtain useful features, whereas Yuan et al ( 42 , 43 ) and Yan et al ( 87 ) incorporated SENet into the ResNet structure to improve tumor segmentation performance.…”

“…Some methods ( 25 , 42 , 43 , 60 , 86 - 88 ) integrate SENet into their own network to improve feature extraction abilities. For example, Qayyum et al ( 60 , 86 ) embedded SENet into the Inception coding part to obtain useful features, whereas Yuan et al ( 42 , 43 ) and Yan et al ( 87 ) incorporated SENet into the ResNet structure to improve tumor segmentation performance. Yousefirizi et al ( 25 ) integrated SENet into the nnU-Net network to improve head and neck tumor segmentation results.…”