Scholarlink

<p>Early and accurate diagnosis of any leukaemia subtype is essential to devise an effective treatment regimen and achieving favourable patient survival outcomes. This paper presents GCSwin-T, a novel Feature Contrastive Graph Distillation Network with leukaemia classifying capabilities, which incorporates contrastive learning, graph-attentive embedding, and feature extraction via a Swin Transformer. The architecture incorporates a cyclic-GAN-based reconstruction module which augments images semantically while preserving the structural cell morphology, and a dual graph neighbourhood scheme which captures spatio-temporal dependencies in contrastive embeddings. Customized T-Blocks within the Swin Transformer enable multi-scale contextual learning via parallel dilated convolutions. The contrastive learning framework enhances feature semantics and class independence based on cosine similarity thresholds, achieving stronger intra-class cohesion and inter-class separability. The GCSwin-T architecture demonstrated an outstanding performance with 98.15% classification accuracy and an F1 score of 98.57% which is better than performance of several state-of-the-art CNN and transformer models. Furthermore, it achieved 100% precision and recall on the clinically critical malignant subtypes of Pro-B leukaemia, validating the clinical reliability of the system. The model alongside detailed ablation studies and explainable AI visualizations has shown the discriminative power and innovative architecture of the system and the model’s relevance to pragmatic workflows in hematopathology GCSwin-T is a scalable, interpretable, and high-fidelity diagnostic system for automated leukemia diagnosis within the next generation system.</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 Comprehensive Study of Various Hybrid Deep Learning Models for Leukaemia Classification: Comparative Analysis with Existing Studies

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Tathagat Banerjee,
  • Davinder Paul Singh,
  • S. Dhanabal,
  • Kothapelli Punnam Chandar,
  • Ravi Kumar Bommisetti,
  • Yogendra Narayan,
  • Nitin Chakole,
  • Chiranjit Dutta,
  • Yogesh H. Bhosale,
  • Ram Murat Singh

摘要

Early and accurate diagnosis of any leukaemia subtype is essential to devise an effective treatment regimen and achieving favourable patient survival outcomes. This paper presents GCSwin-T, a novel Feature Contrastive Graph Distillation Network with leukaemia classifying capabilities, which incorporates contrastive learning, graph-attentive embedding, and feature extraction via a Swin Transformer. The architecture incorporates a cyclic-GAN-based reconstruction module which augments images semantically while preserving the structural cell morphology, and a dual graph neighbourhood scheme which captures spatio-temporal dependencies in contrastive embeddings. Customized T-Blocks within the Swin Transformer enable multi-scale contextual learning via parallel dilated convolutions. The contrastive learning framework enhances feature semantics and class independence based on cosine similarity thresholds, achieving stronger intra-class cohesion and inter-class separability. The GCSwin-T architecture demonstrated an outstanding performance with 98.15% classification accuracy and an F1 score of 98.57% which is better than performance of several state-of-the-art CNN and transformer models. Furthermore, it achieved 100% precision and recall on the clinically critical malignant subtypes of Pro-B leukaemia, validating the clinical reliability of the system. The model alongside detailed ablation studies and explainable AI visualizations has shown the discriminative power and innovative architecture of the system and the model’s relevance to pragmatic workflows in hematopathology GCSwin-T is a scalable, interpretable, and high-fidelity diagnostic system for automated leukemia diagnosis within the next generation system.