PaperlayerView exact revision on arXiv

Source-linked AI summary

Improved Microaneurysm Detection using Deep Neural Networks

Mrinal Haloi

arXiv:1505.04424v2cs.CV

TL;DR

Microaneurysms are an early diabetic-retinopathy lesion, so reliable detection matters for screening. The paper uses a pixel-wise deep neural network with dropout and maxout, without preprocessing or manual feature extraction. It reports better performance than existing methods, including 95% accuracy for Messidor R0 versus R1.

  • Problem

    Microaneurysms usually appear as the first diabetic-retinopathy lesion, making their reliable detection important for screening.

    Source: §I · p. 1

  • Method

    The method classifies fundus-image pixels as MA or non-MA with a deep neural network using dropout and maxout, without preprocessing or manual feature extraction.

    Source: §I · p. 2, §IV · p. 6

  • Results

    The method performs significantly better than existing systems and has 95% accuracy on Messidor for the R0-versus-R1 scenario.

    Source: §III · p. 5

  • Takeaways & Limitations

    The authors describe the system as useful for diabetic-retinopathy mass screening because it achieves state-of-the-art MA-candidate extraction performance with a low false-positive rate.

    Source: §IV · p. 6

    The comparison with existing systems is not on the same ground because datasets and the proportions of images with DR symptoms differ.

    Source: §III · p. 5

Abstract

from arXiv · show

In this work, we propose a novel microaneurysm (MA) detection for early diabetic retinopathy screening using color fundus images. Since MA usually the first lesions to appear as an indicator of diabetic retinopathy, accurate detection of MA is necessary for treatment. Each pixel of the image is classified as either MA or non-MA using a deep neural network with dropout training procedure using maxout activation function. No preprocessing step or manual feature extraction is required. Substantial improvements over standard MA detection method based on the pipeline of preprocessing, feature extraction, classification followed by post processing is achieved. The presented method is evaluated in publicly available Retinopathy Online Challenge (ROC) and Diaretdb1v2 database and achieved state-of-the-art accuracy.

I. INTRODUCTION

Microaneurysms (MAs) are an early diabetic-retinopathy lesion, but fundus-image variation and retinal structures complicate reliable detection. The paper proposes pixel-wise deep learning that requires no preprocessing or manual feature extraction.

I. INTRODUCTION

  • MAs usually appear as the first lesion of diabetic retinopathy, making reliable detection important for screening.In fundus images, they appear as tiny red dots and are described as swollen retinal capillaries.Source: §I · p. 1
  • Vessel crossings, illumination and contrast changes, artifacts, and imaging degradation are among the challenges in fundus-based screening.The introduction identifies Fig. 1 as showing pathological features such as MAs and exudates alongside non-pathological structures such as the optic disc and macula.Source: §I · p. 1, Figure 1 · p. 1
  • Many existing MA-detection systems use a pipeline of preprocessing, manual feature extraction, classification, and postprocessing, often using the green channel.The introduction also describes morphological, filtering-based, and supervised methods using hand-crafted features.Source: §I · p. 1
  • The proposed pixel-wise deep-learning method is described as invariant to luminance, contrast changes, and artifacts, without preprocessing or feature extraction.The authors also state that its performance is independent of vessel structures, the optic disc, and the fovea.Source: §I · p. 2

II. METHODS

The method classifies each image pixel as MA or non-MA from the RGB values in a centered square window.

II. METHODS

  • Each pixel is classified as either MA or non-MA using a deep neural network.The method is motivated by MAs’ gaussian-like intensity distribution and isolated structures.Source: §II · p. 2
  • For each target pixel, the input is a square window of size w centered on that pixel, using all three RGB color channels.The window may also contain other MAs.Source: §II · p. 2
  • The method uses pixel-based classification for the complex task of detecting small microaneurysm structures.Source: §II · p. 2

A. Data Manipulation

The input window is manipulated to suppress neighboring pixels while preserving detail near its center.

A. Data Manipulation

  • The authors introduce foveation and nonuniform sampling as techniques for suppressing neighboring data.This addresses the need to process the rest of a centered window efficiently around local maximum structures of MAs.Source: A. Data Manipulation · p. 2
  • Foveation focuses on the window’s central section and applies increasing Gaussian blur to peripheral pixels as their distance from the center grows.The passage describes this as linear space-invariant Gaussian blur.Source: A. Data Manipulation · p. 2
  • Nonuniform sampling keeps the central window at full resolution while reducing sampling resolution toward the periphery.This allows a large window to be trained using relatively fewer neurons.Source: A. Data Manipulation · p. 2
  • Larger DNN input windows improve performance but increase computational complexity, motivating selective reduction of peripheral pixels.The passage identifies nonuniform sampling as the approach used to make this trade-off.Source: A. Data Manipulation · p. 2

B. Network Architecture

The network combines convolutional and max-pooling layers with fully connected layers and a two-class softmax output.

B. Network Architecture

  • The five-layer network has three convolutional layers, each followed by max-pooling, then a fully connected hidden layer and a softmax classifier.Source: B. Network Architecture · p. 2
  • The softmax layer has two output neurons for MA and non-MA classification.Source: B. Network Architecture · p. 2
  • Dropout training is used for the three convolutional layers and the fully connected hidden layer, while maxout is used in every layer except softmax.Source: B. Network Architecture · p. 2

1) Convolutional Layer:

The convolutional layer is parameterized by input volume size, filter size, depth, stride, and padding; stride should produce an integer output size.

1) Convolutional Layer:

2) Max-Pooling Layer:

Max-pooling selects the maximum within non-overlapping input regions, with the output size determined by input size, spatial extent, and stride. Overlapping pooling is associated with lower overfitting propensity.

2) Max-Pooling Layer:

  • Max-pooling partitions the input into non-overlapping rectangles and selects each subregion's maximum value.Source: 2) Max-Pooling Layer · pp. 2–3
  • Max-pooling is described as providing translation invariance and faster convergence than traditional neural networks.Source: 2) Max-Pooling Layer · p. 2
  • The output volume is calculated from input volume size, spatial extent F, and skipping factor S; when F > S, pooling overlaps and is generally less prone to overfit.Source: 2) Max-Pooling Layer · p. 3

3) Dropout and maxout:

Dropout randomly omits hidden-neuron outputs, while maxout selects the strongest response among groups of units. The paper uses maxout as the activation function in its dropout-trained network.

3) Dropout and maxout:

C. Learning nets

Training uses pixel-centered windows labeled by their center pixel, with rotations and border reflections for data augmentation. Dropout training uses thinned networks, maxout, and max-norm regularization.

C. Learning nets

  • Training samples are windows centered on pixels, labeled MA or non-MA according to expert annotation of the center pixel.Source: C. Learning nets · p. 3
  • Random rotations enlarge the dataset, while horizontal reflections supply pixels beyond image borders and are also used in augmentation.Source: C. Learning nets · p. 3
  • Dropout training drops each neuron with Bernoulli probability p, and forward and backpropagation use only the resulting thinned network.Source: C. Learning nets · p. 3
  • Maxout is associated with improved stochastic-gradient-descent convergence, and max-norm regularization constrains incoming weight norms only in fully connected layers.Source: C. Learning nets · p. 4

III. EXPERIMENTS, RESULTS AND DISCUSSIONS

The detector is evaluated on public datasets and compared with existing screening systems. It reports significantly better performance and a lower false-positive rate, with additional Messidor results and a ROC evaluation limited by unavailable test labels.

III. EXPERIMENTS, RESULTS AND DISCUSSIONS

  • The evaluation uses ROC, Messidor, and Diaretdb1v2; training uses 90,000 MA and 1.5 million non-MA windows with a maxout network and dropout.Each pixel receives six augmented windows; the network architecture is described as three convolutional layers, each followed by max-pooling, a fully connected layer, and a two-neuron softmax output.Source: §III · p. 4, Table I · p. 4
  • At test time, 129 × 129 RGB windows yield pixel-level MA probabilities, and connected regions are filtered by area and convexity.Typical image outputs are reported to reliably detect MA candidates.Source: §III · pp. 4–5
  • The system performs significantly better than existing methods and achieves a lower false-positive rate; Figures 6 and 7 present sensitivity comparisons.The Table II comparison is qualified because datasets and proportions of images with DR symptoms differ.Source: Table II · p. 5, §III · p. 5, Figure 6 · p. 5
  • 95% accuracy, 97% sensitivity, and 94% specificity are reported for Messidor's R0 vs R1 scenario.The comparison is with recent state-of-the-art systems on Messidor.Source: Table III · p. 5, §III · p. 5
  • The ROC evaluation uses part of the training data because test labels are unavailable.Table V compares AUC values with other methods on the same dataset.Source: §III · p. 6, Table V · p. 6

IV. CONCLUSION

The paper presents a five-layer deep network for microaneurysm detection, using dropout training with maxout activation. It reports state-of-the-art candidate-extraction performance with a low false-positive rate and no additional vessel extraction, preprocessing, or feature design.

IV. CONCLUSION

  • The method achieves state-of-the-art MA candidate-extraction performance with a low false-positive rate on publicly available datasets.Source: §IV · p. 6
  • The five-layer network uses dropout training with maxout activation to improve accuracy.Source: §IV · p. 6
  • The system requires no additional blood-vessel extraction, preprocessing, or feature design, and is described as useful for diabetic mass screening.Source: §IV · p. 6
Loading 1505.04424v2…