<p>Accurate blood cell classification is crucial for diagnosing hematological conditions such as anemia, leukemia and infections. Traditional automated methods depend heavily on Fully Supervised Learning (FSL), which requires large volumes of manually labeled data—a resource-intensive and time-consuming process requiring specialized expertise. To overcome these challenges, we propose a novel lightweight Semi-Supervised Active Learning (SSAL) framework that strategically combines Semi-Supervised Learning (SSL) and Active Learning (AL). The motivation behind this integration is rooted in their complementary strengths: SSL utilizes high-confidence predictions to generate pseudo-labels, effectively leveraging vast amounts of unlabeled data, while AL focuses annotation efforts on the most uncertain and informative samples. This synergy ensures that manual labeling is concentrated on data points that mostly improve model performance, significantly reducing the need for large labeled datasets. Experimental results on two publicly available blood cell image datasets, the <i>Blood Cell Count Dataset (BCCD)</i> and the <i>Peripheral Blood Cell (PBC) Dataset</i>, highlight the competitive performance of our SSAL framework compared to state-of-the-art methods, despite using significantly less labeled data. For instance, on the BCCD dataset, our approach achieves an accuracy of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(98.06 \pm 0.34\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>98.06</mn> <mo>±</mo> <mn>0.34</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> with only <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(6\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>6</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of the labeled training data, whereas the state-of-the-art method requires <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(20\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>20</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of labeled data to reach an accuracy of <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(95.13\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>95.13</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>. These results emphasize the potential of integrating SSL and AL to improve both efficiency and accuracy in medical image classification, particularly in resource-constrained scenarios. The code is available at <a href="https://github.com/M-Siyamalan/SSAL">https://github.com/M-Siyamalan/SSAL</a>.</p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

A lightweight semi-supervised active learning framework for blood cell image classification

  • Thedsanamoorthysarma Thanushika,
  • Siyamalan Manivannan,
  • Rajendran Nirthika

摘要

Accurate blood cell classification is crucial for diagnosing hematological conditions such as anemia, leukemia and infections. Traditional automated methods depend heavily on Fully Supervised Learning (FSL), which requires large volumes of manually labeled data—a resource-intensive and time-consuming process requiring specialized expertise. To overcome these challenges, we propose a novel lightweight Semi-Supervised Active Learning (SSAL) framework that strategically combines Semi-Supervised Learning (SSL) and Active Learning (AL). The motivation behind this integration is rooted in their complementary strengths: SSL utilizes high-confidence predictions to generate pseudo-labels, effectively leveraging vast amounts of unlabeled data, while AL focuses annotation efforts on the most uncertain and informative samples. This synergy ensures that manual labeling is concentrated on data points that mostly improve model performance, significantly reducing the need for large labeled datasets. Experimental results on two publicly available blood cell image datasets, the Blood Cell Count Dataset (BCCD) and the Peripheral Blood Cell (PBC) Dataset, highlight the competitive performance of our SSAL framework compared to state-of-the-art methods, despite using significantly less labeled data. For instance, on the BCCD dataset, our approach achieves an accuracy of \(98.06 \pm 0.34\%\) 98.06 ± 0.34 % with only \(6\%\) 6 % of the labeled training data, whereas the state-of-the-art method requires \(20\%\) 20 % of labeled data to reach an accuracy of \(95.13\%\) 95.13 % . These results emphasize the potential of integrating SSL and AL to improve both efficiency and accuracy in medical image classification, particularly in resource-constrained scenarios. The code is available at https://github.com/M-Siyamalan/SSAL.