Retinal Vessel Segmentation Based on the Anam-Net Model

1 Abstract —Accurate of blood vessels can help ophthalmologists diagnose eye-related diseases such as diabetes and hypertension. The task of segmentation of the vessels comes with a number of challenges. Some of the challenges are due to haemorrhages and microaneurysms in fundus imaging, while others are due to the central vessel reflex and low contrast. Encoder-decoder networks have recently achieved excellent performance in retinal vascular segmentation at the trade-off of increased computational complexity. In this work, we use the Anam-Net model to accurately segment retinal vessels at a low computational cost. The Anam-Net model consists of a lightweight convolutional neural network (CNN) along with bottleneck layers in the encoder and decoder stages. Compared to the standard U-Net model and the R2U-Net model, the Anam-Net model has 6.9 times and 10.9 times fewer parameters. We evaluated the Anam-Net model on three open-access datasets: DRIVE, STARE, and CHASE_DB. The results show that the Anam-Net model achieves better segmentation accuracy compared to several state-of-the-art methods. For the DRIVE, STARE, and CHASE DB datasets, the model achieved {sensitivity and accuracy} of {0.8601, 0.9660}, {0.8697, 0.9728}, and {0.8553, 0.9746}, respectively. On the DRIVE, STARE, and CHASE_DB datasets, we also conduct cross-training experiments. The outcome of this experiment demonstrates the generalizability and robustness of the Anam-Net model.


I. INTRODUCTION
Partial and sometimes complete blindness is caused due to the longer duration of chronic eye diseases, including Glaucoma, diabetic retinopathy (DR), and cataracts [1]. These diseases slowly deteriorate different parts of the eye, such as retinal vessels, optic cup (OC), and optic disc (OD). Some people experience substantial effects of these chronic diseases mainly due to their ocular weakness or the severity of the chronic condition. It means that there should be easy access to monitor eye health regularly. The easy and inexpensive availability of regular eye health monitoring procedures will enable us to make a timely diagnosis of the disease, which, in turn, will allow us to avoid the disease or at least delay its impact to a later stage in life. The manual procedures used by physicians/ophthalmologists for the diagnosis of these eye diseases are exhaustive and time-consuming, which sometimes also have difference due to inter-and intraobserver variations. The development of automated methods for the detection of OC, OD, and retinal vessels is of prime importance for the prognosis and diagnosis of Glaucoma and DR [2], [3].
There are two types of glaucoma: closed-angle and openangle. The first type occurs when some parts of the iris of the eye block the flow of fluid, resulting in increased pressure in the eye. Ocular pain, redness of the eye, significant intraocular pressure, and partial vision loss are all symptoms of this kind of Glaucoma. On the other hand, in its second type, the flow of the fluid is not affected. Patients with this type of Glaucoma cannot feel or notice the symptoms until the disease has progressed to an advanced stage, at which point it is irreversible [4].
In manual procedures, the retinal vessels in the retinal fundus images are segmented by the optometrist and ophthalmologists, which is a tiring task and, at the same time, highly susceptible to doctor-to-doctor observationvariation and may also attribute huge discrepancies due to trainer fatigue. As a result, manual segmentation approaches have an inherent limitation on the quality of the extracted data. On the other hand, diagnostics based on computer aided design (CAD) tools is very efficient and can be utilized with high accuracy for the automated diagnosis of chronic eye diseases in a timely manner. Additionally, these automated diagnostics systems based on CAD tools can be used for population-scale screening programs [5].
In retinal fundus images, semantic segmentation is used to detect OC, OD, and retinal vessels. Semantic segmentation is useful for recognizing these constituent parts of the eye, as well as other abnormalities that are needed to characterize chronic eye diseases such as Glaucoma [6], DR [1], [7], age-related macular degeneration (AMD) [8], vascular occlusions [9], and chronic systemic hypoxemia [10].
With recent developments and progress in several fields, including high-performance computing (HPC), machine learning, deep learning, image processing, and machine vision, researchers and academicians have started to devise and explore them for problems in different domains of life. Specifically, several researchers devised deep learning models for pixel-wise semantic segmentation for the detection of OC, OD, and retinal vessels. Most of the developed methods so far characterize huge computational load in addition to attributing lower impact/accuracy without pre-processing. Although some of these previous explorations have achieved better accuracy, very little attention has been paid to the computational complexity of the developed model. This necessitates the development of lightweight deep learning models that have a suitable execution time during both training and testing.
For the diagnosis of DR, retinal vascular segmentation in retinal fundus images must be precise and accurate. Maintaining a reasonable computational complexity for a deep learning (DL) model while aiming for higher accuracy is difficult due to several factors such as tortuosity, higher density, vessel shape, diameter, and the formation of various types of lesions such as microaneurysms, hard exudates, soft exudates, and cotton wool in acquired retinal images [11]. With the progression of chronic eye diseases, there are numerous other issues, such as vessel crossing, centreline reflex, vessel branching, and small vessel formation.
Researchers must take equal consideration to each of the issues described above when developing machine learning and deep learning models for retinal vascular segmentation in retinal fundus images. The previously published convolutional neural network (CNN)-based models in the literature have a higher computational complexity; our goal in this paper is to investigate an alternative model for semantic segmentation of retinal fundus images that requires much less trainable parameters compared to the state-of-theart. The authors in [12] developed the Anam-Net model, which is based on the encoder-decoder architecture. They used Anam-Net for the classification of COVID-19 patients in chest computed tomography (CT) images.
The underlined hypothesis is to investigate and explore a deep learning model, which is deep enough to achieve competitive evaluation metrics compared to best models from the state-of-the-art and at the same time is realizable using normal computing/embedded resources available at the forefront of any healthcare facility. The realization requirement of our hypothesis warrants exploring and developing a deep learning model, which attributes significantly lower trainable parameters and memory overhead.
Hence, the focus of our work is not only on developing deep learning-based automated tool for vessel segmentation, but also to make such tool realizable on ordinary computer systems, which are usually available at forefront in healthcare facility.
The motivation for this work is taken from Anam-Net, which we adapted for vessel detection in retinal fundus images. The fact that it has fewer hyperparameters (resulting in lower computing cost) while preserving competitive evaluation metrics is our primary rationale for adopting it in our study. These attributes make a significant difference when the overall goal is to use the developed model in reallife scenarios in mobile healthcare or point of care in hospitals, where the heavy computating facility is not available. Hence, the attributes of being lightweight and having higher accuracy make the Anam-Net a better and more reliable option to be employed/used at the point of care or mobile health care facility for large-scale screening programs of the governments.
To the best of our knowledge, this is the first attempt to apply Anam-Net for vessel segmentation in retinal fundus images. The contributions are provided below:  We used Anam-Net to segment retinal vessels and were able to achieve competitive results in terms of evaluation metrics while keeping the computational complexity to a minimum.  We chose three publicly available retinal image databases, STARE, DRIVE, and CHASE_DB, to evaluate the performance extensively and to compare the segmentation results to rivals from the state-of-the-art.  We conducted cross-training experiments on the DRIVE, STARE, and CHASE_DB datasets and obtained results that demonstrate the generalization ability and robustness of Anam-Net. The remaining manuscript is organized in the following way. The review of the literature for the segmentation of retinal vessels using machine learning and deep learning models is presented in Section II. The description and implementation details of the basic Anam-Net model are provided in Section III. The results and a detailed comparison of applying the Anam-Net model for retinal vessel segmentation are given in Section IV. Finally, Section V provides conclusions and suggestions for future work.

II. RELATED WORKS
Retinal vessel segmentation is an important task for diagnosing eye diseases, which could be done using manual and automated procedures. Recently, great interest has been shown in applying machine learning and deep learning models for accurate and precise segmentation of the retinal vessels. Accurate segmentation of the retinal vessels helps ophthalmologists reliably diagnose different chronic eye diseases such as Glaucoma and DR. Automated retinal vessel segmentation procedures can be divided into two categories, unsupervised and supervised.
In the literature, different approaches have been adapted for retinal vessel segmentation, such as morphological, kernel-based, vessel tracking, multiscale, model-based, and thresholding [13]. The authors in [14] attempted to apply the kernel-based method to the segmentation of the retinal vessels, where they assumed that the width of the vessels remains constant with distance. This is not a realistic assumption, which limits the kernel-based method for reallife implementation, since both the orientation and width of a retinal vessel are continuously changing. The vessel tracking procedures select a set of starting points to trace the ridges of vessels in images of the retinal fundus [15]. The limitation of the vessel tracking procedure is the user's intervention to follow the ending and starting points. In morphological methods, mathematical equations are used for segmentation of retinal vessels, where mainly top hat operators are applied [16]. The multi-scale detectors were adapted by the authors in [17] for the segmentation of retinal vessels at different orientations and scales. These procedures are efficient, but their performance is degraded for the case of thin vessels. The thresholding-based procedures were implemented by the authors in [18], sometimes leading to an unconnected vascular structure. Model-based approaches perform retinal vessel segmentation, where the basic assumption is to consider the vessels as flexible curves [19]. Model-based approaches are highly sensitive to contrast variations. The interested readers are referred to the extensive review work on unsupervised vessel segmentation performed by the authors in [20].
The authors in [21] developed a model based on a lattice neural network (LNN) for the segmentation of retinal vessels, where the feature extraction was applied as well as reduction. The results were compared to multilayer perceptron (MLP) and support vector machine (SVM) models. They achieved a dice score of 0.66 and 0.69 for the STARE and DRIVE databases.
The authors in [22]- [24] used fully convolutional network (FCN)-based deep neural network (DNN) models for retinal vessel segmentation. However, in the case of [22], there were a large number of false-positive for fundus images having severe pathologies in the optic disc region. On the other hand, the authors in [23] achieved a better segmentation accuracy, but their computational complexity was higher. The authors in [24] also achieved better segmentation accuracy. However, the performance achieved for cross-training results was poor. Specifically, the model trained on the STARE database and tested on images from the DRIVE database produced lower sensitivity, which means lower generalization capability of the developed model.
The authors in [25]- [27] developed and evaluated the different variations of the CNN-based model. The main limitation was the weight gap between the conditional random field (CRF) and CNN, which degraded the overall performance of the developed model. The authors in [28] developed a generative adversarial network (GAN)-based model, where they used a generator along with a patchbased discriminator. The main limitation of the GAN-based models is higher sensitivity to tuning the hyperparameter, higher degree of overfitting, non-convergence, as well as generator gradient vanishing. Due to these different disadvantages, GAN-based models are not suitable for semantic pixel-wise segmentation of retinal fundus images.
Recently, many researchers have targeted retinal vessel segmentation for eye disease classification by developing a different variant of the U-Net model. The authors in [29] developed a Recurrent Residual CNN named "R2U-Net model" for segmentation of retinal vessels. They obtained better dice scores for the images of the CHASE_DB and STARE databases. But they did not validate the generalization capability of their developed model based on the cross-database training and testing strategy. In another attempt, the authors in [30] developed a U-Net-based model, where they performed cross-database training and testing. However, they achieved lower sensitivity for both the case of cross-database training/testing and experiments performed on images of individual datasets. Furthermore, their model attributed a very high computational complexity based on the possession of 36M trainable parameters.
According to [31], a model based on U-Net and deformable convolutional units was proposed. Their proposed model uses a patch size of 48×48 and replaces the original convolutional layer with a deformable convolutional block. The results indicate that their model achieved low sensitivity compared to several state-of-the-art methods. We noticed that other researchers attempted to segment retinal vessels using different variants of the U-Net model [32]- [39]. In most of these studies, the results indicate that breakage of the small retinal vessels occured in the binarization stage, which requires heavy post-processing operations. The heavy post-processing steps in addition to huge computational complexity of these models limits their capability from considering them for point of care deployment in the front end in hospital, which is the main focus of the current exploration.
Due to the advancements in deep learning, medical image analysis is experiencing a paradigm shift [46]. Despite the rapid development and exploration of machine learning and deep learning models for retinal vessel segmentation, numerous problems remained unresolved, which warrant significant consideration for researchers working in the field of medical image segmentation. Some of the previous studies have given some preference to reduce the computational complexity of the model, but we think that there is still much room for improvement. Specifically, earlier explorations did not pay the due attention to handling the memory overhead. The larger memory overhead in addition to the significantly higher computational complexity of the DNN model limits their application from their deployment equipment for large-scale screening programs, which are possible only monitoring the population during their general check-ups at the point of care in the hospital. Additionally, most of the previous explorations avoided the cross-database training and testing experiment, due to which their generalization capability is limited.
We aim to reduce the computational complexity of the deep learning model by reducing the number of trainable parameters. We attempt to reduce the memory overhead of our previously developed model by employing the concept of skip connections after completing max-pooling jobs to share its indices across the encoder and decoder stages [40]. The idea of introducing skip connections and sharing the max-pooling indices helped to improve feature map resolution, which not only reduced computational complexity, but also improved the various evaluation metrics. However, the total number of parameters achieved was more than 8 million, which could be further reduced by looking at current research into deep learning models in other areas of medical image segmentation. Table I lists the methodologies from the literature that we used to compare our segmentation results with.

III. MATERIALS AND METHODS
In this section, first, the fundus images dataset will be discussed along with some pre-processing steps involved in preparing the images for the deep learning model. Next, we will present the Anam-Net architecture and, in the last, we will discuss the implementation details along with the evaluation metrics used to compare the Anam-Net with the state-of-the-art methods.

A. Fundus Image Dataset
The Anam-Net model was tested on three publicly available datasets: DRIVE [47], STARE [14], and CHASE_DB [48]. The DRIVE dataset is made up of 40 fundus images with a resolution of 565×584 pixels that were collected as part of a diabetic retinopathy screening program. The 40 images were separated into two groups, each with 20 images: a training set and a test set. A total of 20 fundus images with a resolution of 605×700 pixels make up the STARE dataset collection. The STARE dataset, unlike the DRIVE dataset, does not include separate training and test data. In this paper, we use a leave-one-out technique for the STARE dataset, in which the model is trained on n-1 samples and evaluated on the remaining one sample. There are 28 images in the CHASE_DB dataset, each with a resolution of 999×960 pixels. A total of 20 images were used to train the network, and the remaining 8 images being used for testing.
The first observer's manual annotations serve as ground truth for our evaluation metrics in all three datasets, which contain two manual segmentations of fundus pictures. Because the image size differs between fundus images in different datasets, we scaled the image to a resolution of 576×576 pixels. We rescale the output probability map to the original size of the test image using bilinear interpolation. The segmentation performance of the Anam-Net is evaluated in the original size of the test image, and thus ensures that the results are not skewed by scale variations.

B. Image Pre-Processing
Pre-processing is essential to suppress noise and improve contrast in the acquired retinal fundus images since they may have non-uniform brightness and intra-and inter-image contrast variations. In this work, we first convert the RGB fundus image to the lightness, channel A, channel B, (LAB) color scheme and apply contrast-limited adaptive histogram equalization (CLAHE) to the lightness channel. The improved L-channel is then combined with the original A and B channels. After that, the image is converted back to the RGB color space, from which the enhanced green channel is retrieved. To further enhance the vessel structure, gamma transformation is used as last pre-processing step.

C. Anam-Net Architecture
The encoder-decoder-based architecture and AD-Block have been combined to make a lightweight CNN-based model. The block diagram of the Anam-Net architecture is shown in Fig. 1. Like the U-Net model, it has a contracting path (encoder) and an expanding path (decoder). With a kernel size of 3 and a filter size of 64, the input fundus image is passed to the convolution layer. A 3×3 convolution is followed by batch normalization and a rectified linear unit (ReLU) layer in each convolution block. A max-pooling layer is incorporated after the convolution block to reduce the dimensionality of the feature by a factor of two. After that, the AD-block is used to learn robust features. The details of the AD-block are shown in Fig. 2. The AD-block comprises of 1×1 depth squeezing convolution, 3×3 feature extraction convolution, and 1×1 depth stretching convolution.
Each encoder and decoder stage in the Anam-Net model has three AD-blocks. In the expansion path, before ADblock, transpose convolution is applied to upsample the feature map to the desired resolution. At the decoder stage, the learnt attributes from the encoder are combined with the layers from the expansion path, enabling the network to learn at multiple levels. For more details on the Anam-Net model, please see the following article [12]. The layer-wise details of the Anam-Net model are shown in Table II.

D. Implementation Details
To train the model, we utilize the Keras deep learning package. With a learning rate of 0.001, a well-known Adam optimizer is utilized. The learning rate is reduced by a factor of 0.1 if the validation loss does not improve after 10 epochs. With a batch size of 4, the model is trained for 150 epochs. To avoid overfitting, we use validation loss as an early stopping criterion. We employ a log dice loss, which focuses on labels that are less accurate [49].   There are only a few training images in each of the three datasets: 20 for DRIVE, 19 for STARE (leave-one-out technique), and 20 for the CHASE_DB. It is very difficult to train a deep learning model with such a short dataset and achieve acceptable segmentation accuracy. We use a variety of data augmentation techniques to improve the robustness and generalization capabilities of the network. Horizontal and vertical flips, random rotations in the range of [0, 360] degrees, random width and height shifts in the range of [0, 0.15], and random magnification in the range of [0.3, 0.12] were among the data augmentation strategies used.
All calculations were performed on IBEX at the King Abdullah University of Sciences and Technology (KAUST) High-Performance Computing (HPC) facility, where we employed a single RTX 2080 Ti GPU for all experiments.

E. Evaluation Metrics
Our proposed model generates a probability prediction map that depicts the likelihood that a pixel belongs to or not to a vessel. For all three datasets, we threshold a probability map with a value of 0.4 to obtain binary segmentation of retinal vessels.
We used standard evaluation metrics that are commonly used in medical image segmentation to evaluate deep learning models. TP, FP, TN, and FN are the abbreviations for true positive, false positive, true negative, and false negative, respectively. Sensitivity (1), specificity (2), accuracy (3), f1-score (5)

TP TN FP FN MCC TP FP TP FN TN FP TN FN
The area under the receiver operating characteristic (ROC) curve (AUC), which ranges from 0 to 1, was used to evaluate image segmentation in addition to the metrics listed above.

IV. RESULTS
In this section, we evaluate the Anam-Net model on the same dataset, i.e., training and test images are both part of the same dataset. Tables III, IV, and V show the results obtained for this experiment. The results highlighted in green, blue, and red indicate the best, second-best, and worst scores of the evaluation metric. As shown in Tables III-V, very few methods have evaluated their proposed model on all evaluation metrics. For this reason, we leave the evaluation metrics empty if it is not evaluated by the respective state-of-the-art method. For comparison, the evaluation metric scores mentioned in Tables III-V of  various methods have been taken from their respective  work. For the DRIVE dataset, the results in Table III show that the Anam-Net model outperforms several methods in terms of sensitivity, accuracy, AUC, and MCC. Our method achieved an f1-score of 0.8152, which is the second-best among all methods listed in Table III. Alom, Yakopcic, Hasan, Taha, and Asari [29] and Wang, Zhao, and Yu [33] achieved the best f1-score and specificity, respectively. However, none of the other evaluation metrics is among the top two, which limits their generalizability. The worst accuracy and sensitivity among all methods is obtained by the models proposed by Zhou, Yu, Xu, Gu, and Yang [25] and Feng, Zhuo, Pan, and Tian [27], respectively.
In the STARE dataset, the Anam-Net model ranked first in terms of sensitivity, f1-score, accuracy, AUC, and MCC, whereas it obtained a competitive score for specificity. Wang, Zhao, and Yu [33] achieved the best specificity and the worst AUC among other methods. Tang, Rui, Yan, Li, and Hu [35] had the second-best specificity and accuracy, whereas the worst sensitivity among the methods listed in Table IV. Regarding the CHASE_DB dataset, the Anam-Net model was ranked first in five out of six evaluation metrics, whereas it ranked second in terms of specificity. In the three datasets, Orlando, Prokofyeva, and Blaschko [22] had the worst f1-score and MCC, while Zhou, Yu, Xu, Gu, and Yang [25] had the worst accuracy.

V. DISCUSSION
In this section, we first perform a cross-database training experiment in which a model trained on one dataset is tested on test images from a second previously unseen dataset. Next, we explore the performance of the Anam-Net model in challenging scenarios such as central vessel reflex and low contrast of fundus images. Finally, we compare the computational complexity of the Anam-Net model with the state-of-the-art methods in terms of the number of trainable parameters.
Table VIII is divided into four sections, where the first section shows the segmentation result on the test images of the STARE dataset when a pre-trained model on the DRIVE dataset is used. The Anam-Net model achieved the best sensitivity, specificity, and accuracy. The method proposed by Mo and Zhang [23] obtained the best f1-score, however, obtained the worst sensitivity among other methods. Oliveira, Pereira, and Silva [24] ranked second in terms of sensitivity, specificity, and accuracy when tested on the STARE dataset. The second section of Table VII shows the segmentation results on the test images of the DRIVE dataset when a pre-trained model on the STARE dataset is used. The Anam-Net ranked first among other methods in three of four evaluation metrics, while ranked second in terms of f1-score. Oliveira, Pereira, and Silva [24] were first in terms of f1-score; however, their method obtained the lowest sensitivity among other methods. When the model is trained on STARE and then tested on DRIVE, we found that it detects fewer thin vessels, resulting in lower sensitivity. On the other hand, the DRIVE database includes substantially more thin vessels annotated than STARE; therefore, sensitivity increased significantly.
To further extend the cross-training experiment, we include the CHASE_DB dataset, which has been reported by very few studies. When the pre-trained model on the DRIVE and STARE datasets is employed, the segmentation results on the test images of the CHASE_DB dataset are shown in sections three and four of Table VIII, respectively. Several methods have been proposed over the past few years to improve the segmentation accuracy of retinal vessels in retinal fundus images. However, very few studies discuss the performance of their deep learning model on challenging scenarios such as central vessel reflex and low contrast. Table IX shows the exemplar results of the Anam-Net in such challenging scenarios. The first column shows the original fundus image with a yellow bounding box, which represents the region of interest for a specific scenario. The second column shows the cropped image of the original fundus image where the image in the top represents central vessel reflex problem, and the image in the bottom represents low contrast. The third and fourth columns show the manual annotations by the first expert and the predicted probability map by the Anam-Net model, respectively. As shown, Anam-Net predicts the retinal vessels with high probability in the central vessel reflex region. For images with low contrast, the Anam-Net model not only predicts the tiny vessels accurately but also segments additional small vessels that were not annotated by the first expert. This experiment validates the robustness of the Anam-Net model for such challenging scenarios.
The Anam-Net model is lightweight with only 4.47M trainable parameters. Only two methods have lower computational complexity compared to Anam-Net, as shown in Table X. Zhuang et al. [39] proposed the Ladder-Net model with only 1.38M parameters. On the CHASE_DB dataset, their model achieved a sensitivity of 0.7978 compared to the Anam-Net with a sensitivity of 0.8553. Moreover, the generalizability of their model is not verified by conducting a cross-training experiment. The results show that the Anam-Net model outperforms state-ofthe-art methods for segmenting the retinal vasculature.
We thoroughly assessed the performance of our developed deep learning model using healthy and pathological images, where we achieved highly compelling results based on several well-known evaluation metrics, including accuracy, specificity, sensitivity AUC, F1-score, and MCC. We have extensively performed comparison with previous best deep learning models from the state-of-the-art, where our achieved results are not only competitive in terms of evaluation metrics, but also its computation complexity is significantly lower. In our previous work in [40], we achieved better performance with almost 8 million trainable parameters. Compared to our previous work and also other best models from the state-of-the-art, our current model achieved improved segmentation performance with almost 4.5 million trainable parameters. Most importantly, our developed deep learning model attributes significantly lower computational complexity and memory overhead based on fewer trainable parameters. These encouraging attributes of the Anam-Net model advocate for its deployment in normal computing resources available at forefront of hospital, to be used for population-scale eye disease screening programs.
Training the Anam-Net takes approximately 1 hour and 20 minutes on a single NVIDIA RTX2080Ti GPU, while it takes about 0.2 seconds to segment a test image of size 576×576 pixels using the trained model. The authors in [23], [29], and [41] reported an average inference time of 0.5 s, 6 s, and 8.4 s, respectively, for a single test image.

VI. CONCLUSIONS
We have developed Anam-Net, which is a lightweight CNN-based encoder-decoder architecture with AD-block as bottleneck layers for semantic segmentation of retinal vessels in fundus images. The performance of our developed model has been thoroughly tested using retinal images from the DRIVE, STARE, and CHASE_DB datasets, where we achieved {sensitivity and accuracy} of {0.8601, 0.9660}, {0.8697, 0.9728}, and {0.8553, 0.9746}, respectively. In all three datasets, the results reveal that the Anam-Net model outperforms the state-of-the-art approaches for segmenting the retinal vessels.
Generally, deep learning models achieve improved performance, but at the expense of high-performance computing (GPUs), which are rarely available at the forefront desk of the healthcare facility. This limits the deep learning-based computer aided tools to be applied in facilities having GPU-based computing resources. The basic hypothesis of this study is to explore and develop a deep learning model which not only achieves competitive semantic segmentation performance in terms of evaluation metrics, but is realizable at the mobile/embedded platforms at the point of care, where the computing resources are not supposed to be GPU-based. Such deep learning models should attribute significantly lower trainable parameters, which will lead to reduced computation complexity and will be realizable at ordinary computing/embedded platforms that are readily available at almost every healthcare facility. In this way, the population-scale screening programs for early diagnosis of vision-threatening eye diseases could be effectively managed.
The developed Anam-Net model has better segmentation accuracy and less trainable parameters compared to several other methods, which make it suitable for easy deployment in mobile platforms at the point of care. On the basis of the achieved results for the cross-database training and testing strategy, we may generalize that the Anam-Net model has significantly higher generalization ability in addition to being lightweight (having significantly lower parameters). Due to attributing significantly lower trainable parameters, the Anam-Net model may be used in a variety of medical imaging applications. Hopefully, in the near future, we will be able to achieve competitive classification performance for the diagnosis of COVID-19 disease using the modified Anam-Net model.