3D ConvNet+: A lightweight adaptive network for 3D medical image segmentation

Recent advancements in 3D medical ViTs, have achieved the best performance on multiple 3D volume constitutes benchmark data for three-dimensional medical image segmentation. The success of these models is attributed to their large receptive fields and extensive model parameters facilitated by non-local self-attention mechanisms. In order to emulate the behavior of large receptive fields exhibited by learning methods, this work introduces a novel lightweight volumetric convolutional network named 3D ConvNet+. Drawing inspiration from the Swin Transformer, the network reexamines volumetric convolutions with large kernel sizes, starting, for instance, from 7×7×7, to achieve a larger global receptive field.Additionally, to better utilize multiscale information, researchers introduce the Lightweight Hybrid Aggregation (LHA) module. LHA leverages adaptive fusion mechanisms with dynamic weight adjustments, providing a crucial advantage to image segmentation structural models with its lightweight and efficient design. We believe that this research methodology will contribute significantly to the academic discourse and practical applications in the field of intelligent segmentation. Furthermore, we anticipate that it will stimulate further in-depth research, fostering advancements in this domain.