Fourier Transform of Percoll Gradients Boosts CNN Classification of Hereditary Hemolytic Anemias

Ario Sadafi, Lucía María Moya Sans, Asya Makhro, Leonid Livshits, Nassir Navab, Anna Bogdanova, Shadi Albarqouni, Carsten Marr

Introduction

Hereditary hemolytic anemias (HHA), including sickle cell disease, thalassemia, and spherocytosis, are genetic disorders characterized by the premature destruction of red blood cells. These conditions often lead to significant alterations in cell shape and density. While genetic testing remains the diagnostic gold standard, its high cost and technical requirements often make it inaccessible in resource-limited or rural clinical settings.

Clinicians have long utilized the Percoll density gradient—a biochemical technique where blood is centrifuged to separate cells by density—as a diagnostic aid. However, interpreting the resulting "band patterns" in the test tubes requires significant expertise. This research explores an automated diagnostic pipeline that combines deep learning with traditional signal processing to classify HHA from Percoll gradient images.

Methodology

The researchers developed a hybrid deep learning architecture designed to extract both spatial and spectral information from images of Percoll gradients. The framework consists of two parallel processing paths:

1. Spatial Path (CNN)

Standard Convolutional Neural Network (CNN) architectures, specifically AlexNet and VGG16, were employed to extract data-driven spatial features. These networks identify visual patterns in the distribution and morphology of the blood bands within the tubes.

2. Spectral Path (FFT)

Recognizing that the distribution of blood cells along the tube can be viewed as a signal of cell density, the team implemented a "handcrafted" spectral path. By smoothing the images and extracting channel intensities, they applied a Fast Fourier Transform (FFT). This process converts the physical distribution of cells into spectral features, capturing the "frequency" and consistency of the bands—a direct proxy for biological markers like cell aggregability.

Fusion Strategies

To determine the most effective way to combine these disparate data types, the study tested two fusion strategies:

  • Early Fusion (EF): Concatenating CNN and FFT features at the first fully connected layer.
  • Late Fusion (LF): Combining features at the second fully connected layer, closer to the final classification output.

Experimental Setup and Dataset

The study utilized a specialized dataset of 143 patients from the Emek Medical Center in Afula. The cohort included:

  • 50 Sickle cell disease cases
  • 35 Thalassemia cases (including alpha and beta subtypes)
  • 11 Spherocytosis cases
  • 47 Healthy controls

Images were captured using standardized holders against white backgrounds to ensure consistency and minimize environmental noise during the feature extraction process.

Results and Analysis

The experimental results demonstrated that the inclusion of spectral features consistently improved classification performance across all models. The AlexNet + Late Fusion model emerged as the top performer, achieving:

  • F1-Score: 0.88 (compared to a 0.86 baseline for AlexNet alone)
  • Accuracy: 0.88 ± 0.01
  • AU ROC: 0.9770 ± 0.0027

Notably, the simpler AlexNet architecture outperformed the more complex VGG16. The authors suggest that VGG16’s higher parameter count likely led to overfitting on the relatively small dataset of 143 samples. However, the addition of FFT features provided a significant boost even to VGG16, raising its AU ROC from 0.9540 to 0.9751.

To ensure model transparency, the team used Grad-CAM (Gradient-weighted Class Activation Mapping). The visualizations revealed that AlexNet tended to focus on the overall consistency of the blood sample, whereas VGG16 focused more specifically on the density and distribution of the bands.

Discussion and Conclusion

This study highlights the power of hybridization in medical imaging, particularly when working with small datasets. While pure CNNs are adept at pattern recognition, the integration of handcrafted spectral features via FFT provides critical inductive biases that improve diagnostic accuracy.

Limitations and Future Directions

As a "proof of concept," the study is limited by its sample size. Future work must involve larger, more diverse datasets to ensure the model generalizes across different populations and clinical environments. Additionally, the model faced challenges with Thalassemia due to the high variability among its subtypes.

Moving forward, the researchers aim to develop standardized clinical protocols and explore whether this automated Percoll analysis can be used to monitor the effectiveness of ongoing therapies, potentially providing a low-cost, automated tool for HHA management in global health settings.

Summary updated

Abstract

Hereditary hemolytic anemias are genetic disorders that affect the shape and density of red blood cells. Genetic tests currently used to diagnose such anemias are expensive and unavailable in the majority of clinical labs. Here, we propose a method for identifying hereditary hemolytic anemias based on a standard biochemistry method, called Percoll gradient, obtained by centrifuging a patient's blood. Our hybrid approach consists on using spatial data-driven features, extracted with a convolutional neural network and spectral handcrafted features obtained from fast Fourier transform. We compare late and early feature fusion with AlexNet and VGG16 architectures. AlexNet with late fusion of spectral features performs better compared to other approaches. We achieved an average F1-score of 88% on different classes suggesting the possibility of diagnosing of hereditary hemolytic anemias from Percoll gradients. Finally, we utilize Grad-CAM to explore the spatial features used for classification.
BibTeX
@inproceedings{sadafi2021fourier,
  author = {Ario Sadafi and Lucía María Moya Sans and Asya Makhro and Leonid Livshits and Nassir Navab and Anna Bogdanova and Shadi Albarqouni and Carsten Marr},
  title = {{Fourier Transform of Percoll Gradients Boosts CNN Classification of Hereditary Hemolytic Anemias}},
  booktitle = {ISBI},
  year = {2021},
  doi = {10.1109/isbi48211.2021.9433788},
}
Figure