Scholarlink

This research work is focused on the automated detection of leukemia using five different models, and leukemia is a fatal blood cancer. It is necessary to detect it early and accurately to save lives, because traditional methods for detecting leukemia are slow, costly and require experts. This research work uses five different models: ConvNext, Xception, InceptionResNet, ViT-b-16, and MobileNet, and it uses 20,000 blood smear images. These im-ages are of five classes: ALL, AML, CLL, CML, and Healthy, thus preprocessing includes resizing, removing noise, enhancing contrast, normalizing, and augmenting data. Previous studies predominantly focused on binary classification (e.g., normal vs. abnormal) or limited subtypes; this work tack-les five distinct classes, addressing the clinical need to differentiate leukemia subtypes. While ViT-b-16 achieved 86.78% accuracy (lower than some prior studies), the increased diagnostic complexity of fine-grained classification enhances clinical utility. ViT-b-16 also shows some overfitting behavior and misclassified some images. Therefore, this study successfully provides an in-sight on how AI could be used to help detect leukemia, so it has the potential to improve the detection process.

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

Performance Evaluating of Deep Learning Models for Leukemia Detection Using Microscopic Blood Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Md. Shakil Hossain,
  • Shafqat Hossain Srijon,
  • Abu Rayhan Akash,
  • Azaz Ahmed,
  • Md Farhad Billah,
  • Md. Tuhid Beg,
  • Md. Parvezur Rahman Mahin,
  • Ahmed Wasif Reza

摘要

This research work is focused on the automated detection of leukemia using five different models, and leukemia is a fatal blood cancer. It is necessary to detect it early and accurately to save lives, because traditional methods for detecting leukemia are slow, costly and require experts. This research work uses five different models: ConvNext, Xception, InceptionResNet, ViT-b-16, and MobileNet, and it uses 20,000 blood smear images. These im-ages are of five classes: ALL, AML, CLL, CML, and Healthy, thus preprocessing includes resizing, removing noise, enhancing contrast, normalizing, and augmenting data. Previous studies predominantly focused on binary classification (e.g., normal vs. abnormal) or limited subtypes; this work tack-les five distinct classes, addressing the clinical need to differentiate leukemia subtypes. While ViT-b-16 achieved 86.78% accuracy (lower than some prior studies), the increased diagnostic complexity of fine-grained classification enhances clinical utility. ViT-b-16 also shows some overfitting behavior and misclassified some images. Therefore, this study successfully provides an in-sight on how AI could be used to help detect leukemia, so it has the potential to improve the detection process.