[Paper Review] A New Multiple Max-pooling Integration Module and Cross Multiscale Deconvolution Network Based on Image Semantic Segmentation

Hongfeng You, Shengwei Tian+4 more|arXiv (Cornell University)|Mar 25, 2020

Radiomics and Machine Learning in Medical Imaging32 references4 citations

TL;DR

This paper proposes MC-Net, a novel deep learning architecture for medical image semantic segmentation that integrates multiple max-pooling features and employs cross multiscale deconvolution to mitigate feature sparsity. By combining multiscale convolutions, a multiple max-pooling integration module, and skip-connected cross-scale deconvolution, MC-Net achieves state-of-the-art performance on binary and multiclass segmentation benchmarks, outperforming FCN, U-Net, PSPNet, and other leading models.

ABSTRACT

To better retain the deep features of an image and solve the sparsity problem of the end-to-end segmentation model, we propose a new deep convolutional network model for medical image pixel segmentation, called MC-Net. The core of this network model consists of four parts, namely, an encoder network, a multiple max-pooling integration module, a cross multiscale deconvolution decoder network and a pixel-level classification layer. In the network structure of the encoder, we use multiscale convolution instead of the traditional single-channel convolution. The multiple max-pooling integration module first integrates the output features of each submodule of the encoder network and reduces the number of parameters by convolution using a kernel size of 1. At the same time, each max-pooling layer (the pooling size of each layer is different) is spliced after each convolution to achieve the translation invariance of the feature maps of each submodule. We use the output feature maps from the multiple max-pooling integration module as the input of the decoder network; the multiscale convolution of each submodule in the decoder network is cross-fused with the feature maps generated by the corresponding multiscale convolution in the encoder network. Using the above feature map processing methods solves the sparsity problem after the max-pooling layer-generating matrix and enhances the robustness of the classification. We compare our proposed model with the well-known Fully Convolutional Networks for Semantic Segmentation (FCNs), DecovNet, PSPNet, U-net, SgeNet and other state-of-the-art segmentation networks such as HyperDenseNet, MS-Dual, Espnetv2, Denseaspp using one binary Kaggle 2018 data science bowl dataset and two multiclass dataset and obtain encouraging experimental results.

Motivation & Objective

• To address the sparsity problem in end-to-end semantic segmentation models, particularly in medical imaging.
• To enhance feature retention and robustness by integrating multi-scale features across encoder and decoder stages.
• To improve segmentation accuracy by reducing parameter count and preserving spatial details through advanced pooling and deconvolution strategies.
• To develop a unified architecture that outperforms existing state-of-the-art models on both binary and multiclass segmentation tasks.

Proposed method

• The encoder uses multiscale convolutions instead of single-channel convolutions to extract hierarchical features.
• A multiple max-pooling integration module fuses features from different encoder submodules using 1x1 convolutions and applies max-pooling with varying kernel sizes to enhance translation invariance.
• The decoder employs cross multiscale deconvolution, where each deconvolution submodule is fused with corresponding multiscale features from the encoder to recover spatial resolution and context.
• The network uses a pixel-level classification layer for final segmentation prediction.
• The model is trained end-to-end using standard cross-entropy loss on medical image datasets.
• The integration of multi-scale pooling and skip-connected deconvolution reduces feature sparsity and improves feature representation.

Experimental results

Research questions

• RQ1 Can a multiple max-pooling integration module effectively reduce feature sparsity while preserving spatial context in semantic segmentation?
• RQ2 How does cross multiscale deconvolution improve feature recovery and segmentation accuracy compared to standard deconvolution?
• RQ3 To what extent does the proposed MC-Net architecture outperform existing state-of-the-art models on medical image segmentation benchmarks?
• RQ4 Does the use of multiscale convolutions in both encoder and decoder stages enhance feature representation and robustness?

Key findings

• MC-Net achieved state-of-the-art performance on the binary Kaggle 2018 Data Science Bowl dataset, outperforming FCN, U-Net, and PSPNet.
• On multiclass medical image segmentation datasets, MC-Net demonstrated superior segmentation accuracy compared to HyperDenseNet, MS-Dual, ESPNetv2, and DenseASPP.
• The multiple max-pooling integration module effectively reduced the number of parameters while maintaining high feature representation quality.
• The cross multiscale deconvolution strategy significantly improved feature recovery and reduced sparsity after max-pooling operations.
• The model achieved consistent performance gains across diverse medical imaging tasks, indicating strong generalization capability.