[Paper Review] MiniSeg: An Extremely Minimum Network for Efficient COVID-19 Segmentation

Yu Qiu, Yun Liu+2 more|arXiv (Cornell University)|Apr 21, 2020

COVID-19 diagnosis using AIMedicine64 references43 citations

TL;DR

MiniSeg is an 83K-parameter lightweight network using an Attentive Hierarchical Spatial Pyramid (AHSP) for fast, data-efficient segmentation of COVID-19 infected areas in chest CT slices, achieving competitive accuracy with high efficiency and easy retraining.

ABSTRACT

The rapid spread of the new pandemic, i.e., COVID-19, has severely threatened global health. Deep-learning-based computer-aided screening, e.g., COVID-19 infected CT area segmentation, has attracted much attention. However, the publicly available COVID-19 training data are limited, easily causing overfitting for traditional deep learning methods that are usually data-hungry with millions of parameters. On the other hand, fast training/testing and low computational cost are also necessary for quick deployment and development of COVID-19 screening systems, but traditional deep learning methods are usually computationally intensive. To address the above problems, we propose MiniSeg, a lightweight deep learning model for efficient COVID-19 segmentation. Compared with traditional segmentation methods, MiniSeg has several significant strengths: i) it only has 83K parameters and is thus not easy to overfit; ii) it has high computational efficiency and is thus convenient for practical deployment; iii) it can be fast retrained by other users using their private COVID-19 data for further improving performance. In addition, we build a comprehensive COVID-19 segmentation benchmark for comparing MiniSeg to traditional methods.

Motivation & Objective

• Address the challenge of COVID-19 CT segmentation with limited labeled data and the need for fast training/inference.
• Develop a lightweight yet accurate segmentation model to reduce overfitting and computational cost.
• Propose the AHSP module to enable effective multi-scale feature learning in a compact network.
• Create and benchmark a COVID-19 segmentation dataset suite to compare MiniSeg against state-of-the-art methods.

Proposed method

• Introduce AHSP: multi-branch dilated depthwise separable convolutions with exponential dilation rates for multi-scale learning.
• Hierarchical feature fusion across scales with an attention mechanism to highlight relevant structures while suppressing noise.
• Two-path encoder with Downsampler Blocks and nested skip connections to enhance multi-scale encoding.
• Decoder with Feature Fusion Modules and progressive upsampling to recover fine details, plus deep supervision.
• Efficient parameter usage: K=4 branches and grouped operations to reduce parameters and FLOPs, achieving 83K parameters and high speed.
• Training with data augmentation on four public COVID-19 CT segmentation datasets; 80 epochs with Adam and poly learning rate schedule.

Experimental results

Research questions

• RQ1 Can an extremely lightweight network achieve competitive segmentation accuracy on COVID-19 CT data with limited training samples?
• RQ2 Does the AHSP module improve multi-scale feature learning without inflating model size?
• RQ3 How does MiniSeg compare to state-of-the-art segmentation methods in accuracy, speed, and parameter efficiency on public COVID-19 CT datasets?

Key findings

• MiniSeg achieves 83K parameters, about two orders of magnitude smaller than many baselines, with high inference speed.
• AHSP-enabled multi-scale learning improves segmentation performance in a compact model.
• On multiple datasets, MiniSeg provides competitive mIoU, sensitivity, and Dice scores while maintaining high specificity and fast speed.
• For COVID-19-CT100: mIoU 82.15, SEN 84.95, SPC 97.72, DSC 75.91, HD 74.42.
• For COVID-19-P9: mIoU 85.31, SEN 90.60, SPC 99.15, DSC 80.06, HD 58.46.
• For COVID-19-P20: mIoU 84.49, SEN 85.06, SPC 99.05, DSC 76.27, HD 51.06.
• For COVID-19-P1110: mIoU 78.33, SEN 79.62, SPC 97.71, DSC 64.84, HD 71.69.
• MiniSeg delivers up to 516.3 fps, illustrating exceptional efficiency.