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

TL;DR

This paper presents a method to detect sickle cell disease using deep learning techniques that work well with limited data and provide explainable results.

Contribution

The paper introduces a novel combination of transfer-learning, contrastive-learning, and XAI for SCD detection in low-resource settings.

Findings

Models using transfer learning and triplet loss outperformed traditional loss functions.

Explainable AI methods were integrated to enhance model transparency in clinical applications.

The approach is effective for SCD detection with limited training data.

Abstract

Sickle cell disease (SCD) is a severe hereditary blood disorder that affects millions worldwide, necessitating early and accurate detection to improve patient outcomes. State-of-the-art approaches for automatic detection of SCD use deep learning (DL) based models, which require a large amount of training data for efficient training. However, such large training datasets are often not available, significantly limiting the efficiency of DL-based models. In this paper, we propose different approaches to address this issue. Firstly, we propose to use a transfer-learning based approach, where we use pre-trained models like ResNet-50, DenseNet-121, and EfficientNet-B0 and fine-tune them for SCD detection. To further enhance the efficiency of the models, we then propose to include contrastive-learning-based approach using triplet loss. We also use focal loss to handle class imbalance. Additionally, we integrate Explainable Artificial Intelligence (XAI) methodologies to interpret and explain the model’s predictions, ensuring transparency and trustworthiness in clinical settings. Experiments on a publicly available SCD image dataset show that models trained with transfer learning and triplet loss outperform those trained with binary cross-entropy or focal loss.

Figures