Abstract:In recent years, advanced U-like networks have demonstrated remarkable performance in medical image segmentation tasks. However, their drawbacks, including excessive parameters, high computational complexity, and slow inference speed, pose challenges for practical implementation in scenarios with limited computational resources. Existing lightweight U-like networks have alleviated some of these problems, but they often have pre-designed structures and consist of inseparable modules, limiting their application scenarios. In this paper, we propose three plug-and-play decoders by employing different discretization methods of the neural memory Ordinary Differential Equations (nmODEs). These decoders integrate features at various levels of abstraction by processing information from skip connections and performing numerical operations on upward path. Through experiments on the PH2, ISIC2017, and ISIC2018 datasets, we embed these decoders into different U-like networks, demonstrating their effectiveness in significantly reducing the number of parameters and FLOPs while maintaining performance. In summary, the proposed discretized nmODEs decoders are capable of reducing the number of parameters by about 20% ~ 50% and FLOPs by up to 74%, while possessing the potential to adapt to all U-like networks. Our code is available at <a class="link-external link-https" href="https://github.com/nayutayuki/Lightweight-nmODE-Decoders-For-U-like-networks" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve several key problems in U - like networks in medical image segmentation tasks: 1. **Excessive parameters**: Existing U - like networks usually contain a large number of parameters, which leads to an increase in model complexity and training difficulty. 2. **High computational complexity**: These networks require high - computing resources, limiting their deployment in practical application scenarios, especially in cases where computing resources are limited. 3. **Slow inference speed**: Due to the increase in parameters and computational complexity, the inference speed of these networks is slow, affecting the performance of real - time applications. 4. **Poor adaptability**: Although existing lightweight U - like networks have alleviated the above problems to a certain extent, they are usually pre - designed structures and their modules are not separable, limiting their application in different scenarios. To solve these problems, the authors propose a new method - simplifying the decoder part of the U - like network by introducing the neural memory Ordinary Differential Equations (nmODEs) decoder. Specifically, they propose three different discretization methods (explicit Euler method, Heun method, and linear multi - step method) and verify the effectiveness of these methods through experiments. ### Main contributions - **Reduction in parameters and FLOPs**: By using the nmODEs decoder, the number of parameters (about 20% - 50%) and the number of floating - point operations (up to 74%) can be significantly reduced while maintaining performance. - **Wide applicability**: The proposed decoder can be adapted to all U - like networks, improving the universality and flexibility of the model. - **Performance improvement**: In some cases, using the nmODEs decoder can not only reduce parameters and computational amount, but also slightly improve the performance of the model. ### Experimental verification The authors conducted experiments on the PH2, ISIC2017, and ISIC2018 datasets, and the results show that: - **Significant reduction in parameters and FLOPs**: For example, after applying EED, the parameters of Att - UNet are reduced by about 46% and the FLOPs are reduced by about 74%; after applying HD, the parameters of UNet are reduced by about 30% and the FLOPs are reduced by about 40%. - **Performance maintained or improved**: Most networks maintain or improve their performance after integrating the nmODEs decoder. In particular, on the PH2 dataset, Att - UNet and UNet reach the state - of - the - art levels in the mIoU and DSC indicators respectively. In conclusion, this paper proposes an effective method to solve the problems of excessive parameters, high computational complexity, and slow inference speed in U - like networks in medical image segmentation tasks, and shows its superior performance on multiple datasets.

A Lightweight U-like Network Utilizing Neural Memory Ordinary Differential Equations for Slimming the Decoder

M-Net: A Lightweight Network Based on Multilayer Perceptron for Massive MIMO CSI Feedback

CFPNet-M: A Light-Weight Encoder-Decoder Based Network for Multimodal Biomedical Image Real-Time Segmentation

REQ-YOLO

A Holistically-Guided Decoder for Deep Representation Learning with Applications to Semantic Segmentation and Object Detection.

A Lightweight Encoder–Decoder Path for Deep Residual Networks

More complex encoder is not all you need

Optimizing Medical Image Segmentation with Advanced Decoder Design

Towards bi-directional skip connections in encoder-decoder architectures and beyond

Lightweight Low-Power U-Net Architecture for Semantic Segmentation

LMANet: A Lightweight Asymmetric Semantic Segmentation Network Based on Multi-Scale Feature Extraction

Neural Memory State Space Models for Medical Image Segmentation

An Ultra-Efficient Memristor-Based DNN Framework with Structured Weight Pruning and Quantization Using ADMM

LEDNet: A Lightweight Encoder-Decoder Network for Real-Time Semantic Segmentation

Decoders Matter for Semantic Segmentation: Data-Dependent Decoding Enables Flexible Feature Aggregation

Boosted Neural Decoders: Achieving Extreme Reliability of LDPC Codes for 6G Networks

1M parameters are enough? A lightweight CNN-based model for medical image segmentation

Bilateral attention decoder: A lightweight decoder for real-time semantic segmentation

Deep Decoder: Concise Image Representations from Untrained Non-convolutional Networks

LFC-UNet: learned lossless medical image fast compression with U-Net

UDC: Unified DNAS for Compressible TinyML Models