Fig. 2

Key components of the PUResNetV2.0 architecture. a The overall architecture, highlighting the integration of an encoder path for input downsampling and a decoder path for upsampling the feature maps, with skip connections between the corresponding blocks in both paths. b Illustration of the convolution block used within the encoder for input downsampling and feature extraction, which is composed of Minkowski convolutional layer, batch normalization layer, and ReLU activation function. c Presentation of the transpose block, which is deployed in the decoder path for input upsampling and consists of a Minkowski convolution transpose layer, a Minkowski batch normalization layer, and a Minkowski ReLU activation function. d Depiction of the ResNet-inspired basic block, which possesses skip connections for effective feature extraction and is utilized in both the encoder and decoder paths