TL;DR
This paper introduces MSK&MP, a novel texture descriptor for OCT images, designed to accurately distinguish between various ocular diseases like AMD and CNV, even in the presence of noise. The method achieves impressive classification accuracies, outperforming current state-of-the-art deep learning techniques on challenging datasets.
What Problem Does This Paper Address?
Age-related macular degeneration (AMD) and its neovascular form, choroidal neovascularization (CNV), are leading causes of blindness. Early and accurate diagnosis is crucial for effective treatment. Optical Coherence Tomography (OCT) is a key imaging modality for visualizing the posterior eye, but analyzing these images, especially when they are small or degraded by noise (like speckle noise), can be challenging for ophthalmologists. This paper aims to develop an automated system to precisely detect, classify, and diagnose these conditions by creating a robust feature extraction method that can handle noisy OCT images.
Methods
The core of this work is the development of a new texture descriptor called Multi-Size Kernels Echo-Weighted Median Patterns (MSK&MP). This method is an extension of Local Binary Patterns (LBP) and aims to overcome LBP's limitations, particularly its susceptibility to noise.
Here's a breakdown of the MSK&MP approach:
- Preprocessing: Standard OCT image preprocessing steps are applied, including removing white areas, resizing, binarization using Otsu's thresholding, applying a non-weighted median filter to remove noise, morphological dilation to close fluid regions, and flattening retinal layers using polynomial fitting.
- Feature Extraction (MSK&MP):
- Echo-Weighted Median Calculation: Instead of directly using pixel values, MSK&MP calculates a median value (
Imedian) from a kernel. Crucially, it uses "echo-weights" derived from a Gaussian function to determine how many times each pixel value within the kernel is repeated. This process is designed to give more weight to the center pixel and reduce the impact of noisy pixels.
- Echo-Weighted Median Calculation: Instead of directly using pixel values, MSK&MP calculates a median value (
- Multi-Size Kernels: The method employs multiple kernel sizes (e.g., 3x3, 5x5, 7x7, 3x5, 5x7, 3x7). This is because different kernel sizes can capture varying textural details, and combining them provides a more comprehensive representation of the image's texture.
- Median LBP Calculation: For each kernel size, a modified LBP is calculated using the echo-weighted median. This involves comparing surrounding pixel values to the
Imedianand encoding the results into a binary pattern. - Feature Vector Generation: Histograms are generated from these multi-size MSK&MP images, and these histograms are concatenated to form the final feature vector for each OCT image.
- Feature Weighting, Selection, and Classification:
- PCA: Principal Component Analysis (PCA) is used for dimensionality reduction.
- NCA: Neighborhood Component Analysis (NCA) is employed for feature weighting, assigning importance scores to features.
- Iterative Feature Dropping: Features with low weights are iteratively dropped to optimize performance and prevent overfitting.
- Classifiers: Several classifiers are tested, including Support Vector Machines (SVM) with polynomial and Gaussian kernels, Random Forest, Naïve Bayes, Adaboost, and RUS-Boost.
Datasets
The proposed MSK&MP method was evaluated on three distinct datasets:
- Dataset 1 (Duke University, Harvard University, University of Michigan): This dataset consists of SD-OCT volumetric scans from 45 patients, categorized into 15 normal, 15 dry AMD, and 15 Diabetic Macular Edema (DME).
- Dataset 2 (Noor Eye Hospital): This dataset comprises 148 SD-OCT volumes, with 48 AMD, 50 DME, and 50 Normal cases.
- Dataset 3 (Mendeley database): This dataset contains 12,641 OCT images, including Normal, Drusen (early AMD), and CNV cases. The authors specifically selected "worst-case condition images" for their experiments.
Key Insights and Ideas
- Echo-Weighting for Noise Robustness: The core innovation of MSK&MP is the "echo-weighting" mechanism. By repeating pixel values based on Gaussian-derived weights, the method effectively smooths out the impact of individual noisy pixels on the median calculation, leading to more robust texture features.
- Multi-Size Kernels Capture Diverse Textures: The use of multiple kernel sizes is crucial. Different sizes are adept at capturing different scales of texture, from fine details to broader regional patterns. Combining features from these diverse kernels provides a richer representation for classification.
- Hand-Crafted Features Can Compete with Deep Learning: Despite the rise of deep learning, this work demonstrates that well-designed hand-crafted features like MSK&MP can achieve competitive, and in some cases superior, performance, especially when dealing with noisy or limited datasets where deep learning models might struggle with overfitting or require extensive training.
- Feature Weighting and Selection are Critical: The paper highlights the importance of not just extracting features but also intelligently selecting and weighting them. PCA and NCA play a significant role in refining the feature set and improving classification accuracy.
Key Results
- High Accuracy on Multiple Datasets: MSK&MP achieved an impressive overall accuracy of 99.78% on Dataset 1 (Duke University) and 96.59% on Dataset 2 (Noor Eye Hospital) using SVM with a polynomial kernel.
- Superior Performance: The method outperformed current state-of-the-art deep learning techniques and vision transformers, as well as other texture descriptors.
- Excellent Binary Classification: For binary classifications (e.g., Normal vs. AMD, Normal vs. DME), MSK&MP achieved perfect or near-perfect accuracy (100%) in several instances.
- Robustness to Noise: The paper specifically highlights the performance on the "noisy Noor Eye Hospital dataset," where it achieved over 96% accuracy, demonstrating its resilience to image degradation.
- Ablation Study Insights: An ablation study showed that combining features from multiple kernel sizes (5x5, 7x7, 3x5, 5x7, and 3x7) yielded the best results, indicating the synergistic benefit of diverse kernel sizes.
Limitations and Future Work
- Limitations:
- The paper notes that using very large kernel sizes (e.g., up to 3x9 or 5x9) did not yield further improvements and was omitted from results, suggesting a potential trade-off between kernel size and performance gains.
- While the method is robust to speckle noise, the paper doesn't explicitly detail its performance on other types of noise or artifacts.
- Future Work:
- The authors suggest extending the work to 3D retinal images by extracting volume and depth information.
- They propose using these 3D structures for multi-view classification, which could provide a more comprehensive understanding of disease spread and progression.
Summary updated