Hybrid Deep Learning Framework for Anemia Detection and Classification Using RBC Images and CBC Data

Technical Summary: Hybrid Deep Learning Framework for Anemia Detection and Classification

Problem Statement
Anemia, characterized by a reduction in red blood cells (RBCs) or hemoglobin levels, is a prevalent hematological disorder affecting significant portions of the global population, particularly women, children, and pregnant women. While routine clinical diagnostics such as Complete Blood Count (CBC) analysis and peripheral blood smear examinations are effective, they are often time-consuming, expensive, and require trained personnel. The paper addresses the need for rapid, inexpensive, and automated screening procedures that can assist in early detection and severity grading, especially in resource-limited settings.

Methodology
The authors propose a hybrid framework that integrates two distinct data modalities: morphological analysis of RBC smear images and quantitative analysis of CBC report data.

1. Image-Based Classification (RBC Smear Analysis):

   • Dataset: The study utilizes a dataset of segmented peripheral blood smear images categorized into healthy and anemic classes.
   • Preprocessing: Images are resized to 224x224 pixels, normalized to handle intensity variations, and augmented (rotation, flipping, zooming, brightness adjustment) to mitigate overfitting.
   • Model Comparison: Five pre-trained Convolutional Neural Network (CNN) architectures were evaluated using transfer learning over 30 epochs: ResNet-18, VGG19, MobileNetV2, EfficientNet-B0, and DenseNet121.
   • Selection: DenseNet121 was selected as the optimal architecture due to its dense connection mechanism, which facilitates feature reuse, improves gradient flow, and reduces vanishing gradients. It demonstrated superior capability in learning fine-grained RBC features.

2. CBC-Based Severity Prediction:

   • Data: The system utilizes standard hematological parameters including Hemoglobin (HGB), Hematocrit (HCT), Mean Corpuscular Volume (MCV), Mean Corpuscular Hemoglobin (MCH), MCHC, RBC count, WBC count, Platelet count (PLT), and Red Cell Distribution Width (RDW).
   • Modeling: A Stacking Ensemble approach was developed to predict anemia severity.
     • Base Classifiers: K-Nearest Neighbors (KNN), Random Forest (RF), and XGBoost.
     • Meta-Classifier: Logistic Regression.
   • Process: Base classifiers are trained independently, their predictions are aggregated as new features, and the meta-classifier generates the final severity prediction.

Key Contributions

• Hybrid Architecture: The paper presents a dual-pathway system that combines the visual diagnostic power of deep learning (for morphology) with the statistical predictive power of ensemble machine learning (for clinical parameters).
• Architectural Benchmarking: A comparative analysis of five state-of-the-art CNNs specifically for RBC smear classification, identifying DenseNet121 as the most effective for this specific medical imaging task.
• Ensemble Optimization: The development of a stacking classifier for CBC data, demonstrating that combining diverse algorithms (KNN, RF, XGBoost) yields higher accuracy than individual models for structured hematological data.

Results

• Image Classification: DenseNet121 achieved the highest performance among the tested CNNs, reaching 94.0% accuracy on the test set. The model demonstrated a precision, recall, and F1-score of 0.94, with a validation accuracy of 95.0%. The inference time was recorded at 0.2585 seconds per image on a CPU, suggesting suitability for practical triage.
• Severity Prediction: The stacking ensemble model applied to CBC data achieved 98.0% accuracy in predicting the severity of anemia. The confusion matrix indicated high reliability in distinguishing between different severity levels.
• Generalization: Training and validation curves for the DenseNet121 model showed minimal divergence (loss decreasing from 0.1325 to 0.0798), indicating strong generalization capabilities without significant overfitting.

Significance and Claims
The authors claim that this framework offers a "thorough and practically useful solution" for anemia monitoring and grading. By integrating strong image analysis capabilities with machine learning for CBC analysis, the system aims to improve efficiency in routine workflows. The paper posits that the proposed method is effective and deployable for rapid screening, capable of assisting healthcare professionals in early detection and symptom severity assessment. The study concludes that the hybrid approach successfully meets the objectives of automatic and efficient diagnosis based on both visual and clinical characteristics.

Limitations and Future Directions
The authors acknowledge that while the framework is promising, future work requires validation with larger and more complex clinical samples. They also suggest the integration of explainable AI (XAI) techniques to enhance clinical interpretability and the development of new hematological parameters. The ultimate goal mentioned is the direct deployment of the system in diagnostic settings.