Abstract
Sickle Cell Disease (SCD) is a genetic disorder characterized by unpredictable clinical severity. Traditional assessment of disease severity relies on expensive laboratory measurements of hemoglobin density, which are often unavailable in low-resource settings. This paper presents SCD-severity-GCN, a deep learning pipeline that predicts SCD severity using accessible data: images of Percoll density gradients and basic ultrasound measurements. By leveraging Graph Convolutional Networks (GCNs) to model patient similarities and a hybrid feature extraction method, the proposed model achieves performance comparable to methods utilizing high-end laboratory equipment.
Introduction
Sickle Cell Disease (SCD) causes red blood cells (RBCs) to become rigid and sickle-shaped, leading to severe complications such as vaso-occlusive crises and organ damage. A key biomarker for severity is the presence of hypochromic (low hemoglobin) and hyperchromic (dense) cells. However, quantifying these cells requires automated cell counters and specialized tests that are scarce in high-prevalence, low-resource regions.
Percoll density gradients offer a cost-effective alternative by separating cells by density into visible bands. However, manual interpretation of these gradients is subjective and difficult. This research aims to automate this process, providing a reliable severity score (1-4) using only a centrifuge, a camera, and basic clinical data like spleen size.
Methods
The authors developed a multi-stage pipeline, SCD-severity-GCN, to translate Percoll images into clinical predictions.
Feature Extraction and Proxy Estimation
To capture the unique visual characteristics of Percoll gradients, the model employs a hybrid CNN-FFT approach. A standard CNN (AlexNet) extracts textural features, while a Fast Fourier Transform (FFT) captures the periodic banding patterns of cells. Crucially, to maintain independence from expensive lab tests, a separate CNN was trained to estimate the percentage of hypo- and hyperchromic cells directly from the images.
Population Graph and GCN
Instead of processing samples in isolation, the model constructs a population graph where each node represents a patient sample. Edges are weighted based on similarity in spleen size (obtained via ultrasound) and the estimated hemoglobin density. A Graph Convolutional Network (GCN) then processes this graph, allowing the model to refine predictions by considering the clinical profiles of similar patients.
Results
The study utilized 216 samples from 17 patients diagnosed with SCD, with ground-truth severity determined by the Sebastiani scale. The SCD-severity-GCN achieved an accuracy of 0.61, significantly outperforming traditional machine learning baselines like Support Vector Machines (accuracy < 0.45).
Performance Benchmarking
A key finding was the small performance gap between the image-only model and an "upper bound" model (GCN-Lab) that used actual laboratory measurements (accuracy 0.65). This suggests that the proxy estimation of hemoglobin density from images is highly effective, making the system viable for clinics without high-end lab equipment.
Feature Disentanglement
Visualization of the high-dimensional features using UMAP embeddings demonstrates the GCN's ability to cluster samples by severity. While raw image features appear scattered, the GCN-processed features show clear separation between severity classes, indicating successful feature disentanglement.
Discussion and Conclusion
The proposed method demonstrates that deep learning can bridge the gap between low-cost diagnostic tools and high-end laboratory analysis. By automating the interpretation of Percoll gradients and incorporating clinical similarity through graph networks, the system provides a robust tool for SCD severity assessment.
The primary limitation of this study is the small patient cohort (17 patients), which may lead to overfitting. Future work will focus on expanding the dataset to include more diverse cases, particularly pediatric data, and refining the severity scale toward a continuous prediction model. This approach holds significant promise for improving SCD management in underserved regions.
Summary updated