Page 235 - Computational Retinal Image Analysis

232 CHAPTER 12 Diabetic retinopathy and maculopathy lesions

4.5 Deep learning
Deep learning (DL) algorithms use hierarchical models with multiple levels of
transformations to learn highly abstract representations of data. They revolutionized
the world of machine learning and were becoming increasingly popular in medical
image analysis. Unfortunately, publicly available retinal imaging datasets are scarce,
small, and lack necessary annotations, which severely limit DL applicability to this
domain. Nevertheless, authors believe that this will change in the forthcoming future,
and we will see more DL method used for lesion detection. As such we decided to
create a separate category for DL algorithms, even though they are machine learning-
based methods.
Convolutional neural networks (CNNs) are the main DL algorithm to deal
with image data because they are able to learn the multidimensional relationships
between data points (pixels/voxels). van Grinsven et al. [46] presented a selective
data sampling technique to improve the CNN training time. They combined it
with a vanilla-type CNN to find HMs in color fundus images. Orlando et al. [47]
used CNN as a feature extractor and combined it with the random forest classifier
to find MAs. Chudzik et al. [48] created a novel CNN architecture with inception
modules to segment exudates. They showed that transfer knowledge between even
small retinal datasets of different modalities results in better performance. Next, they
proposed a novel fine-tuning scheme called “interleaved freezing” that optimizes
the transfer learning process [49]. They combined it with a U-Net type CNN to find
MAs. Subsequently, they improved their results by combining their dedicated CNN
with an auxiliary codebook built from the network’s intermediate layers output
[50]. The auxiliary codebook allowed to identify the most difficult data samples and
improved their segmentation results. Finally, they proposed a fully CNN with batch
normalization layers and DICE loss function to segment MAs [20]. Fig. 9 depicts
an example of MA detection in FPs. Compared to other methods that require the
aforementioned five stages of image analysis, the proposed algorithm required only
two: preprocessing and classification. Dai et al. [51] proposed a clinical report guided
multisieving CNN which combined multimodal information from text reports with
image data. Clinical reports are used to bridge the semantic gap between low-level
image features and high-level diagnostic information.

4.6 Miscellaneous

Miscellaneous techniques are all remaining works that do not fit in previous
categories. Agurto et al. [52] proposed a multiscale amplitude-modulation-
frequency-modulation (AM-FM) method to find both BL and RL. The cumulative
distribution functions of instantaneous value of frequency, relative instantaneous
value of frequency angle, and instantaneous value of amplitude were used for
feature analysis. Javidi et al. [53] proposed a technique which is used 2D Morlet
wavelet to find MA candidates. At the next stage, a discriminative dictionary
learning approach was employed to distinguish MAs from other structures.
Quellec et al. [54] detected lesions by locally matching a lesion template in sub-