[Paper Review] Semi-Supervised Active Learning for COVID-19 Lung Ultrasound Multi-symptom Classification

Lei Liu, Wentao Lei+4 more|arXiv (Cornell University)|Sep 9, 2020

COVID-19 diagnosis using AI27 references4 citations

TL;DR

This paper proposes a semi-supervised Two-Stream Active Learning (TSAL) framework for multi-symptom classification of COVID-19 lung ultrasound images, leveraging label correlation via a multi-label margin strategy and confidence validation to reduce annotation costs. The method achieves state-of-the-art performance using only 14.7% of the full training data, validated on a new large-scale dataset (COVID19-LUSMS) with 6,836 images from 71 patients.

ABSTRACT

Ultrasound (US) is a non-invasive yet effective medical diagnostic imaging technique for the COVID-19 global pandemic. However, due to complex feature behaviors and expensive annotations of US images, it is difficult to apply Artificial Intelligence (AI) assisting approaches for lung's multi-symptom (multi-label) classification. To overcome these difficulties, we propose a novel semi-supervised Two-Stream Active Learning (TSAL) method to model complicated features and reduce labeling costs in an iterative procedure. The core component of TSAL is the multi-label learning mechanism, in which label correlations information is used to design multi-label margin (MLM) strategy and confidence validation for automatically selecting informative samples and confident labels. On this basis, a multi-symptom multi-label (MSML) classification network is proposed to learn discriminative features of lung symptoms, and a human-machine interaction is exploited to confirm the final annotations that are used to fine-tune MSML with progressively labeled data. Moreover, a novel lung US dataset named COVID19-LUSMS is built, currently containing 71 clinical patients with 6,836 images sampled from 678 videos. Experimental evaluations show that TSAL using only 20% data can achieve superior performance to the baseline and the state-of-the-art. Qualitatively, visualization of both attention map and sample distribution confirms the good consistency with the clinic knowledge.

Motivation & Objective

• To address the high cost and complexity of annotating lung ultrasound images for multiple COVID-19 symptoms in clinical settings.
• To reduce labeling effort in multi-symptom, multi-label classification of lung US images using active learning and semi-supervised learning.
• To develop a model that leverages label correlations to improve sample selection and pseudo-label confidence in low-data regimes.
• To build and release a new large-scale, multi-label annotated lung US dataset (COVID19-LUSMS) for COVID-19 symptom classification.
• To validate that model predictions align with clinical expertise through explainable attention visualization.

Proposed method

• Proposes a Two-Stream Active Learning (TSAL) framework with a sample stream for selecting informative images and a label stream for assigning confident pseudo-labels.
• Employs a multi-label margin (MLM) strategy that uses label correlation to prioritize samples with high uncertainty and high potential information gain.
• Introduces a confidence validation mechanism to filter and assign reliable pseudo-labels, reducing reliance on human annotation.
• Uses a multi-symptom multi-label (MSML) classification network to learn discriminative features from lung US images across multiple symptoms.
• Incorporates human-machine interaction (HMI) to confirm final annotations, enabling iterative model fine-tuning with minimal human labeling.
• Leverages Grad-CAM for explainability, visualizing attention maps to validate alignment with clinical diagnostic knowledge.

Experimental results

Research questions

• RQ1 Can a semi-supervised active learning framework reduce labeling costs in multi-symptom lung ultrasound classification without sacrificing performance?
• RQ2 How effective is label correlation modeling via the multi-label margin (MLM) strategy in selecting informative samples for active learning?
• RQ3 To what extent can confidence validation of pseudo-labels reduce human annotation effort in the active learning loop?
• RQ4 Can the proposed TSAL framework achieve performance comparable to full-supervision using less than 20% of the labeled data?
• RQ5 Do attention maps from the trained model align with clinical expertise in identifying pathological regions in lung US images?

Key findings

• The TSAL framework achieves superior classification performance using only 14.7% of the full training data, outperforming baseline and state-of-the-art methods.
• The multi-label margin (MLM) strategy achieves the highest accuracy and stability across all active learning iterations, with the best final performance at 20th iteration.
• Specificity for B-line and pleural effusion improved significantly—from 81.81% to 92.49%—demonstrating effective handling of class imbalance.
• Human annotation effort was reduced to nearly zero after the first iteration when confidence validation was applied, showing high efficiency of the label stream.
• Grad-CAM visualizations confirmed strong consistency between model attention and clinical diagnosis, particularly for A-line and B-line features.
• The proposed method achieved near or better performance than full-supervision baselines using only 16.6% to 27.6% of the labeled data across different strategies.