TL;DR
This paper investigates the reliability of using retinal vascular fractal dimension (FD) as a biomarker for diabetic retinopathy detection. The authors found that FD measurements are highly sensitive to variations in image quality, segmentation methods, and regions of interest, leading to significant errors and making it an unreliable biomarker for clinical applications without strict standardization.
What Problem Does This Paper Address?
Diabetic retinopathy (DR) is a leading cause of blindness, and early detection is crucial for effective treatment. The complex branching pattern of retinal blood vessels, quantifiable by fractal dimension (FD), has been proposed as a potential biomarker for DR. However, existing research presents conflicting findings regarding the association between FD and DR, suggesting a lack of robustness in its measurement. This paper aims to systematically investigate the factors contributing to this variability and assess the reliability of FD as a clinical biomarker. The core problem is that while FD theoretically captures vascular complexity, its practical application is hampered by inconsistencies in how it's measured, making it difficult to trust its diagnostic value.
Methods
The study employs a comprehensive pipeline to calculate fractal dimensions from retinal images. This pipeline involves several key steps:
- Image Preprocessing: Local normalization of luminosity and contrast to handle variations in image illumination.
- Vessel Segmentation: This is a critical step. The paper evaluates three state-of-the-art automatic methods (Frangi's vesselness, Soares', and Zhang's) for RGB images and a specialized method (BIMSO) for SLO images. These methods produce "soft" (probability) segmentation maps, which are then converted to "hard" (binary) segmentations using a threshold.
- Region of Interest (ROI) Determination: Automatic detection and segmentation of the optic disc (OD) and fovea are performed to define circular ROIs of varying sizes (4x, 5x, and 6x the OD radius) around the fovea.
- Fractal Dimension Calculation: Three classic FD calculation methods are used: Box Dimension ($D_B$), Information Dimension ($D_I$), and Correlation Dimension ($D_C$). These methods quantify different aspects of the fractal nature of the vascular network.
The study then systematically analyzes the impact of variations in:
- Manual Annotations: Comparing FD values from two human observers.
- Automatic Segmentation Methods: Comparing FD from different algorithms.
- Regions of Interest: Evaluating FD within different sized circular areas.
- Segmentation Accuracy: Investigating how varying segmentation thresholds affect FD.
- Imaging Modalities: Comparing FD from images captured by different fundus cameras and an SLO camera.
Datasets
The study utilizes three datasets:
- MESSIDOR Database: A large dataset (1200 images) of color fundus images with diabetic retinopathy grades (R0-R3), used for investigating intergroup and intragroup FD variations.
- DRIVE Database: Contains 40 fovea-centered color retinal images from 33 subjects, with manual annotations by two ophthalmologists. This dataset was used for studying the stability and robustness of FD measurements, particularly concerning manual annotations and segmentation methods.
- 5 Cameras Dataset: A newly established dataset comprising images from 12 healthy volunteers captured by five different fundus cameras (3nethra Classic, Canon CR-1 Mark II, Nidek AFC-230, Topcon NW300, and EasyScan SLO). This dataset was specifically created to investigate variations due to different imaging modalities and cameras.
Key Insights and Ideas
- FD is highly sensitive: The core takeaway is that FD measurements are not robust. Even small changes in image processing, segmentation, or the area analyzed can lead to significant variations in the calculated FD.
- Manual annotations are not perfect: While considered "ground truth," even expert human annotations can differ, particularly in their inclusion of fine vessels, which impacts FD.
- Automatic segmentation is a major source of error: The choice of automatic segmentation algorithm and its accuracy (determined by thresholding) dramatically influences FD values, often leading to results that are not well-correlated with manual segmentations.
- Region of interest matters: The specific area of the retina analyzed (e.g., full image vs. a smaller circle around the fovea) significantly impacts FD, suggesting that a standardized ROI is crucial.
- Imaging modality and camera type introduce variability: Different cameras capture images with varying resolutions, illumination, and noise levels, leading to inconsistent FD measurements for the same subject.
- FD is a "deceptive" biomarker: Due to these sensitivities, the authors conclude that FD, as currently measured, is not a reliable biomarker for quantitative clinical applications without stringent standardization.
Key Results
- Inter-observer variability: Maximum relative errors between two observers were up to 7.1% for $D_B$, with mean relative errors around 1.9%.
- Segmentation method variability: Maximum relative errors for automatic methods compared to manual segmentation were around 9.3% for $D_B$, with mean relative errors around 4.3%. Pearson correlation tests indicated weak associations between automatic methods and manual ground truth.
- Region of interest variability: Using different ROIs (4x, 5x, 6x OD radius) resulted in relative errors up to 3.8% for $D_B$, though these were significantly associated (p < 0.01).
- Segmentation accuracy impact: A segmentation accuracy (MCC) of 78% yielded a mean relative error of 10.42% in FD, while lower accuracies (e.g., 63.74%) led to higher errors (14.44%).
- Camera variability: Different cameras showed mean relative errors of FD up to 2.80% for repeated acquisitions of the same subject.
- Intra-observer repeatability: Across 5 cameras, the average variation for repeated acquisitions on the same subject was 1.11%.
Limitations and Future Work
Limitations:
- Sensitivity to parameters: The study highlights the extreme sensitivity of FD to various parameters, including segmentation thresholds and ROI definitions.
- Conflicting literature: The paper acknowledges that previous studies have reported conflicting results, which this work aims to explain by demonstrating the instability of FD measurement.
- Current clinical applicability: The findings suggest that FD is not yet ready for direct clinical use without significant methodological improvements and standardization.
Future Work:
- Direct measurement from probability maps: The authors suggest that future work should explore measuring FD directly from vessel probability maps, rather than binary segmentations, to potentially mitigate errors introduced by thresholding.
- Standardization: The need for standardized protocols for image acquisition, segmentation, and ROI selection is strongly implied.
- Investigating other biomarkers: While FD is the focus, the broader challenge of finding reliable biomarkers for DR detection remains.
Summary updated