Hybrid multi-model framework with knowledge distillation for precise and efficient sickle cell detection.

Accurate and timely detection of sickle red blood cells has a significant impact on the patients' treatment. Developing a computer vision system that is both lightweight and resource-efficient is important for enabling real-time edge-based applications in resource-constrained healthcare environments. This study proposes a hybrid sickle cell detection framework that uses Bi-lightweight Knowledge Distillation (KD)-guided CNNs, combining two lightweight student models from different teacher networks with Machine Learning (ML) classifiers to identify sickle cells with accuracy and efficiency. Using depthwise separable convolution layers, lightweight models are created that drastically lower model complexity. These lightweight models are improved by extracting knowledge from top-performing teacher networks, such as MobileNetV2 and DenseNet121. Using this Knowledge Distillation learning technique, lightweight models can enhance their representational capabilities without adding to their computational complexity. To improve decision-making performance, the Bi-distilled student models are combined with conventional machine learning KNN classifiers to create a hybrid ensemble that combines the advantages of ML-based classification with dual KD-guided CNN-based feature extraction. The Sickle Cell Disease Dataset is used to train and validate the proposed system. The learned student models are then converted into TensorFlow Lite (TFLite) format using Float32 quantization to further reduce computational complexity. In combination with dual KD-guided CNN models and a KNN classifier, the proposed hybrid model achieves a high classification accuracy of 95.45% and an F1-score of 95.02% while being small at only 3.96 MB. Deploying and evaluating the framework on a resource-constrained Raspberry Pi 4 Model B demonstrate its potential for real-time, edge-based, and point-of-care applications.