Dhar Prasenjit, Suganya Devi K, Bhattacharjee Ramanuj, Srinivasan P
Medical Imaging Laboratory, Department of Computer Science and Engineering, National Institute of Technology Silchar, Silchar, Assam, India.
Institute of Technical Education and Research, Department of Computer Science and Engineering, Siksha O Anusandhan University, Bhubaneswar, Odisha, India.
Microsc Res Tech. 2025 May;88(5):1566-1581. doi: 10.1002/jemt.24786. Epub 2025 Jan 27.
Red blood cells (RBCs) or Erythrocytes are essential components of the human body and they transport oxygen from the lungs to the body's tissues, regulate balance, and support the immune system. Abnormalities in RBC shapes (Poikilocytosis) and sizes (Anisocytosis) can impede oxygen-carrying capacity, leading to conditions such as anemia, thalassemia, McLeod Syndrome, liver disease, and so on. Hematologists typically spend considerable time manually examining RBC's shapes and sizes using a microscope and it is time-consuming. The proposed LSTM based neural network (NN) deep-learning strategy helps to classify abnormal RBCs automatically and accurately and overcome blood-related disorders at an early stage. After data processing, traditional and high-level features are fused to clearly distinguish between abnormal RBC classes. Class imbalance favors the dominant class, resulting in biased forecasts. To address class imbalance, a custom loss function is generated by integrating class weights and loss functions before feeding fused features to the NN classifier. Specifically, the loss function is designed to assign higher penalties to the misclassification of underrepresented classes, ensuring that the model is more sensitive to these classes during training. This is achieved by integrating class weights directly into the cross-entropy loss calculation, thereby balancing the influence of each class on the model's learning process. The proposed approach's performance is evaluated using the publicly accessible Chula-PIC-Lab dataset and privately gathered dataset from the Cachar Cancer Hospital and Research Centre (CCHRC) in Assam, India. The proposed approach achieves an average of and -score and accuracy on the Chula-PIC-Lab dataset and an average of and -score and accuracy on the CCHRC dataset for and classes and surpasses benchmark models including Custom CNN, Custom LSTM, Efficient Net-B1, SMOTE, Hybrid NN, and HPKNN.
红细胞（RBCs）或红血球是人体的重要组成部分，它们将氧气从肺部输送到身体组织，调节平衡，并支持免疫系统。红细胞形状（异形红细胞症）和大小（红细胞大小不均）的异常会阻碍氧气携带能力，导致贫血、地中海贫血、麦克劳德综合征、肝病等病症。血液学家通常花费大量时间使用显微镜手动检查红细胞的形状和大小，这很耗时。所提出的基于长短期记忆网络（LSTM）的神经网络（NN）深度学习策略有助于自动、准确地对异常红细胞进行分类，并在早期阶段克服与血液相关的疾病。经过数据处理后，传统特征和高级特征相融合，以清晰区分异常红细胞类别。类别不平衡有利于主导类别，导致预测有偏差。为了解决类别不平衡问题，在将融合特征输入到神经网络分类器之前，通过整合类别权重和损失函数生成自定义损失函数。具体而言，损失函数旨在对代表性不足的类别的误分类施加更高的惩罚，确保模型在训练期间对这些类别更敏感。这是通过将类别权重直接整合到交叉熵损失计算中来实现的，从而平衡每个类别对模型学习过程的影响。使用公开可用的朱拉 - PIC - 实验室数据集以及从印度阿萨姆邦卡恰尔癌症医院和研究中心（CCHRC）私下收集的数据集对所提出方法的性能进行评估。对于 和 类别，所提出的方法在朱拉 - PIC - 实验室数据集上平均达到 和 分数以及准确率，在CCHRC数据集上平均达到 和 分数以及准确率，并超过了包括自定义卷积神经网络（Custom CNN）、自定义长短期记忆网络（Custom LSTM）、高效网络 - B1（Efficient Net - B1）、合成少数过采样技术（SMOTE）、混合神经网络（Hybrid NN）和高维核近邻网络（HPKNN）在内的基准模型。
