Sickle cell disease detection in low-resource conditions using transfer-learning and contrastive-learning coupled...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Genetics and Genomics

Background:

• Sickle cell disease (SCD) is a prevalent hereditary blood disorder impacting millions globally.
• Accurate and early detection of SCD is crucial for improving patient prognosis and outcomes.
• Current deep learning (DL) models for automated SCD detection require extensive training datasets, which are often unavailable.

Purpose of the Study:

• To address the challenge of limited training data for deep learning-based SCD detection.
• To enhance the efficiency and accuracy of automated SCD detection models.
• To integrate Explainable Artificial Intelligence (XAI) for transparent and trustworthy clinical decision support.

Main Methods:

• Utilized transfer learning by fine-tuning pre-trained models (ResNet-50, DenseNet-121, EfficientNet-B0) for SCD detection.
• Incorporated contrastive learning with triplet loss to further boost model performance.
• Employed focal loss to manage class imbalance within the dataset.
• Integrated Explainable Artificial Intelligence (XAI) techniques for model interpretability.

Main Results:

• Models employing transfer learning and triplet loss demonstrated superior performance compared to those trained with binary cross-entropy or focal loss alone.
• The proposed methods effectively addressed the limitations posed by small training datasets.
• XAI integration provided insights into model predictions, enhancing clinical trust.

Conclusions:

• Transfer learning combined with contrastive learning offers a robust solution for SCD detection with limited data.
• The developed models show promise for accurate, efficient, and transparent automated SCD diagnosis.
• Integrating XAI is vital for the clinical adoption of AI-driven diagnostic tools in healthcare.