Hierarchical 3D Feature Learning forPancreas Segmentation

Radio astronomical images object detection and segmentation: A benchmark on deep learning methods

“…U-Net++ [49] proposes a combination of architectural improvements by improving the skip connection pathways and extending the deep supervision mechanism, while Tiramisu [50] employs DenseNets [51] instead of ResNets [34] improving the performance using fewer parameters. Other approaches [11,52] propose hierarchical decoding for segmentation, which, instead of employing the pipeline of downsampling and upsampling path, decode features at multiple resolutions and combine them at the end.…”

“…Then the image is expanded back to the original image size in the decoding path to obtain the binary masks that classify the objects. There are many variants that extend this framework by applying skip connections [47], deep supervision [48], or hierarchical decoding [11].…”

“…In our analysis, we employ the following architectures on our radio-astronomical dataset, as shown in Fig. 5: a basic encoder-decoder architecture, U-Net [47], which adds skip connections to the basic architecture, U-Net with deep supervision, which enhances the loss by computing it also on intermediate upsampled masks, Tiramisu [50], which employs DenseNets [51] instead of convolutions in the downsampling and upsampling blocks, and PankNet [11], which makes use of hierarchical decoding for information mixing between upsampling paths.…”

“…Finally, we employ PankNet [11], which does not make use of the downsampling-upsampling pipeline but applies encoders and decoders in a multilevel way, as shown in Fig 5(d). Each decoder receives only the corresponding encoder output as input, while the other versions concatenated it to the previous decoder output.…”

“…To overcome these limitations, deep learning models represent the evolution of such approaches and yield interesting results extensively explored in several domains. In particular, object detection and semantic segmentation models based on deep learning are currently used in different domains, such as automotive [5][6][7][8], medical imaging [9][10][11], video surveillance [12,13], and robot navigation [14][15][16]. The radio-astronomical domain has not yet been exhaustively explored with the application of the mentioned methods; therefore, this work represents an attempt to gather performance and computational requirements about several state-of-the-art approaches to be used as a reference for future work.…”