Scholarlink

<p>Accurate identification of malignant blood cells is essential for early diagnosis and treatment planning in patients with hematologic diseases. A dual-fusion multi-representation deep learning framework is proposed for interpretable and robust classification of blood cells. Rather than combining independent acquisition modalities, the proposed model integrates seven diagnostically complementary image-derived representations extracted from the same underlying RGB blood-smear image, including raw appearance, stain-deconvolved, pseudo-3D, wavelet, texture, morphological, and graph-based representations, through attention-weighted mid-level fusion and Choquet-integral-based decision fusion. Experiments were performed on a multiclass dataset of curated data of five cell types: myeloblasts, erythroblasts, monocytes, basophils and segmented neutrophils. The overall multiclass accuracy of the model was 97.2%, its macro F1 was 97.2%, and for binary malignancy detection, the model's accuracy was 98.8%, with an AUC-ROC of 0.993, a sensitivity of 98.0%, a specificity of 99.3%, and an F1 score of 98.5%. Calibration evaluation revealed an expected calibration error (ECE) of 1.3% and a Brier score of 0.036, indicating well-calibrated probabilistic outputs within the evaluated dataset setting. The Choquet interaction analysis revealed the highest synergy between the M3 (pseudo-3D) and M6 (morphology) representation branches (+ 0.19), suggesting biologically consistent fusion behavior between multiscale structural and morphometric cues. Robustness to Gaussian noise (σ = 0.05) and 30% occlusion held &gt; 95% macro-F1, indicating resistance to imaging disturbances. Computationally, the model took 65 epochs to train on an NVIDIA RTX A5000 GPU (24&#xa0;GB VRAM) with mixed precision under PyTorch 2.1, using only 2.6&#xa0;GB of memory at inference and averaging ≈47&#xa0;ms per image latency, allowing for near real-time screening. The results support the proposed framework as a promising, computationally efficient, and interpretable multi-representation approach for blood-cell image analysis, while broader multi-institutional and clinically annotated validation remains necessary before clinical deployment.</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

Dual-fusion multi-representation deep learning framework for interpretable blood cell classification and malignancy detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Eshika Jain,
  • Vinay Kukreja,
  • Pratham Kaushik,
  • Malvinder Singh Bali,
  • Ambuj Kumar Agarwal,
  • Anuj Kumar Jain

摘要

Accurate identification of malignant blood cells is essential for early diagnosis and treatment planning in patients with hematologic diseases. A dual-fusion multi-representation deep learning framework is proposed for interpretable and robust classification of blood cells. Rather than combining independent acquisition modalities, the proposed model integrates seven diagnostically complementary image-derived representations extracted from the same underlying RGB blood-smear image, including raw appearance, stain-deconvolved, pseudo-3D, wavelet, texture, morphological, and graph-based representations, through attention-weighted mid-level fusion and Choquet-integral-based decision fusion. Experiments were performed on a multiclass dataset of curated data of five cell types: myeloblasts, erythroblasts, monocytes, basophils and segmented neutrophils. The overall multiclass accuracy of the model was 97.2%, its macro F1 was 97.2%, and for binary malignancy detection, the model's accuracy was 98.8%, with an AUC-ROC of 0.993, a sensitivity of 98.0%, a specificity of 99.3%, and an F1 score of 98.5%. Calibration evaluation revealed an expected calibration error (ECE) of 1.3% and a Brier score of 0.036, indicating well-calibrated probabilistic outputs within the evaluated dataset setting. The Choquet interaction analysis revealed the highest synergy between the M3 (pseudo-3D) and M6 (morphology) representation branches (+ 0.19), suggesting biologically consistent fusion behavior between multiscale structural and morphometric cues. Robustness to Gaussian noise (σ = 0.05) and 30% occlusion held > 95% macro-F1, indicating resistance to imaging disturbances. Computationally, the model took 65 epochs to train on an NVIDIA RTX A5000 GPU (24 GB VRAM) with mixed precision under PyTorch 2.1, using only 2.6 GB of memory at inference and averaging ≈47 ms per image latency, allowing for near real-time screening. The results support the proposed framework as a promising, computationally efficient, and interpretable multi-representation approach for blood-cell image analysis, while broader multi-institutional and clinically annotated validation remains necessary before clinical deployment.