[論文レビュー] MDU-Net: Multi-scale Densely Connected U-Net for biomedical image segmentation

Jiawei Zhang, Yuzhen Jin+3 more|arXiv (Cornell University)|Dec 2, 2018

AI in cancer detection参考文献 62被引用数 73

ひとこと要約

MDU-Net は U-Net に三つのマルチスケール密結合（エンコーダ、デコーダ、クロスブロック）を追加し、マルチスケール特徴を融合することで、より深いネットワークと GlaS データセットにおける腺セグメンテーションの改善を実現する。量子化は過剰適合を抑えつつ、精度を維持または向上させる。

ABSTRACT

Biomedical image segmentation plays a central role in quantitative analysis, clinical diagnosis, and medical intervention. In the light of the fully convolutional networks (FCN) and U-Net, deep convolutional networks (DNNs) have made significant contributions to biomedical image segmentation applications. In this paper, we propose three different multi-scale dense connections (MDC) for the encoder, the decoder of U-shaped architectures, and across them. Based on three dense connections, we propose a multi-scale densely connected U-Net (MDU-Net) for biomedical image segmentation. MDU-Net directly fuses the neighboring feature maps with different scales from both higher layers and lower layers to strengthen feature propagation in the current layer. Multi-scale dense connections, which contain shorter connections between layers close to the input and output, also make a much deeper U-Net possible. Besides, we introduce quantization to alleviate the potential overfitting in dense connections, and further improve the segmentation performance. We evaluate our proposed model on the MICCAI 2015 Gland Segmentation (GlaS) dataset. The three MDC improve U-Net performance by up to 1.8% on test A and 3.5% on test B in the MICCAI Gland dataset. Meanwhile, the MDU-Net with quantization obviously improves the segmentation performance of original U-Net.

研究の動機と目的

• Motivate improved biomedical image segmentation by enhancing information flow in U-Net through multi-scale dense connections.
• Propose three dense connectivity patterns (encoder, decoder, cross-connection) to fuse high-level and low-level features across scales.
• Incorporate network quantization to reduce overfitting while preserving segmentation accuracy.
• Evaluate on MICCAI 2015 GlaS gland segmentation dataset to quantify gains over standard U-Net.

提案手法

• Introduce three multi-scale dense connected blocks: dense encoder, dense decoder, and dense cross connections.
• Fuse feature maps across neighboring scales by down-sampling and up-sampling to the same resolution before concatenation.
• Adopt 1x1 convolutions to control channel numbers and maintain a small parameter overhead.
• Evaluate different configurations (Min/Mout, upper/lower cross connections) to study the impact of dense connectivity.
• Apply Incremental Quantization (INQ) to parameter weights to reduce overfitting and assess performance with partial quantization.

実験結果

リサーチクエスチョン

• RQ1 Does integrating multi-scale dense connections in encoder, decoder, and cross-blocks improve segmentation accuracy over standard U-Net?
• RQ2 Which dense connectivity pattern (encoder, decoder, cross) yields the best performance, and how does their combination affect accuracy and overfitting?
• RQ3 Can quantization further mitigate overfitting in densely connected U-Nets without sacrificing accuracy on biomedical segmentation tasks?
• RQ4 How do the proposed methods perform on a standard gland segmentation benchmark (GlaS) compared to U-Net and related architectures?

主な発見

• Three multi-scale dense connections improve segmentation performance over U-Net, achieving up to 3% higher mean Dice on Test A and up to 4.1% on Test B when combined.
• Individually, dense encoder, dense decoder, and dense cross blocks each yield gains over U-Net across evaluation metrics and data splits.
• The combination of three dense connections provides the best results on average for Test A, with up to 3%/4.1% improvements reported when compared to U-Net.
• Quantization via Incremental Quantization (INQ) helps reduce overfitting and can maintain or improve Dice scores, with half-quantized configurations showing competitive performance (e.g., improvements on Test B).
• The proposed MDU-Net maintains a modest parameter increase relative to U-Net, while delivering improved segmentation accuracy, suggesting it can serve as a robust backbone for U-shaped biomedical segmentation.