A SYSTEM AND METHOD FOR RETINAL VEIN OCCLUSION (RVO) DETECTION

Abstract

ABSTRACT A SYSTEM AND METHOD FOR RETINAL VEIN OCCLUSION (RVO) DETECTION The present disclosure provides a system (100) and method (200) for retinal vein occlusion (RVO) detection, comprising a plurality of distributed client nodes (110), each including a local dataset (112) of retinal images categorized into healthy and RVO-affected classes, and a local processing module (114) configured to train a convolutional neural network (CNN). A central aggregation server (120) coordinates the training across the client nodes (110) using a federated learning module (122). The central aggregation server (120) also includes a technique selection module (124) configured to dynamically switch between a federated averaging (FedAvg) technique (126) for independent and identically distributed (IID) data and a stochastic controlled averaging technique (128) for non-IID data. This system (100) enhances model accuracy by adapting to varying data distributions while preserving data privacy and minimizing communication overhead.

Specification

Description:FIELD
The present disclosure relates to the healthcare domain and specifically focuses on systems and methods for detecting retinal vein occlusion (RVO).
DEFINITION
As used in the present disclosure, the following terms are generally intended to have the meaning as set forth below, except to the extent that the context in which they are used indicates otherwise.
• Federated Learning: The term "Federated Learning" refers to a machine learning approach where multiple institutions or clients train their local models on their own data without sharing it with a central server. Instead, only the model updates are shared and combined to create a global model, ensuring that sensitive data remains private while still contributing to a more accurate, collective learning process.
• Client Drift: The term "client drift" refers to the phenomenon where individual local models trained on data from different clients (hospitals or institutions) diverge significantly from each other or from the global model due to differences in data distribution. This divergence makes it challenging to combine these local models effectively, which can slow down the learning process and reduce the accuracy of the global model.
• Global Model: The term "global model" refers to the combined machine learning model that is built by aggregating updates from all the local models trained by the clients. In the context of federated learning, this global model benefits from the knowledge gained from multiple data sources while maintaining data privacy and security.
• Convergence Speed: The term "convergence speed" refers to the rate at which the global model in the federated learning system reaches its optimal performance or accuracy. Faster convergence means that the model learns to make accurate predictions more quickly, which is crucial for timely and effective decision-making, especially in medical applications like retinal disease detection.
• Homomorphic Encryption: The term "homomorphic encryption" refers to a technique used to secure data during the training process by encrypting model updates. This allows computations to be performed on the encrypted data without needing to decrypt it, ensuring that sensitive information remains protected even while being used to improve the global model.
• Federated Averaging (FedAvg) Technique: The term "Federated Averaging (FedAvg) technique" refers to a method used in federated learning to aggregate model updates from different clients by averaging them. This technique is particularly effective when the data across the clients is similar (IID), as it combines the updates in a way that enhances the global model's accuracy without requiring access to the raw data.
• Stochastic Controlled Averaging technique: The term "Stochastic Controlled Averaging technique" refers to an advanced federated learning method designed to address the challenges of training models on non-independent and identically distributed (non-IID) data. In federated learning, variations in data distributions across clients can lead to inconsistencies in model updates, commonly known as client drift. The technique mitigates this issue by using control variates, which are correction terms applied to the local model updates. These control variates guide the local models to stay closer to the global model during training, thereby stabilizing the updates and improving convergence speed. The technique ensures that even when clients have significantly different data distributions, the global model can still learn effectively from the aggregated information, leading to faster convergence and higher accuracy.
• Non-IID Data: The term "non-IID" refers to non-independent and identically distributed data, meaning that the data across clients in the federated learning system is different or unevenly distributed. Non-IID data can cause variations in the local models, leading to challenges in creating a unified global model due to the diverse nature of the training data.
• Differential Privacy Mechanism: The term "differential privacy mechanism" refers to a technique used to add noise to the model updates before they are shared with the central server, ensuring that no single data point from the client's dataset can be traced or identified. This approach provides an additional layer of privacy protection, making it difficult for attackers to infer sensitive information from the model updates.
• Gradient Updates: The term "gradient updates" refers to the changes or adjustments that each client makes to its local model during training. These updates are based on the errors identified in the model's predictions and are used to guide the model towards more accurate predictions. In federated learning, these gradient updates are shared with the central server instead of the actual data.
• RVO-Affected Eyes: The term "RVO-affected eyes" refers to the condition where the veins in the retina are blocked, leading to vision problems. In the context of the disclosure, the system is designed to detect and differentiate between healthy eyes and those affected by Retinal Vein Occlusion (RVO) using advanced machine learning models.
The above definitions are in addition to those expressed in the art.
BACKGROUND
The background information herein below relates to the present disclosure but is not necessarily prior art.
Machine learning and AI have significantly advanced data analysis in medical imaging, aiding in the early diagnosis of conditions like retinal vein occlusion (RVO). Centralized data collection for these models raises privacy concerns, prompting a shift toward federated learning, which enables collaborative model training without sharing raw data.
Traditional centralized machine learning systems compromise data security by pooling sensitive data from multiple institutions. Federated learning improves privacy by training models locally and sharing only updates, but it struggles with non-uniform (non-IID) data across clients, slowing convergence and reducing model accuracy.
Current approaches face issues with data privacy risks, client drift in non-IID scenarios, high communication costs, computational constraints on client nodes, and limited adaptability to real-time data changes.
Therefore, there is a need for a system and method for retinal vein occlusion (RVO) detection that alleviates the aforementioned drawbacks.
OBJECTS
Some of the objects of the present disclosure, which at least one embodiment herein satisfies, are as follows:
It is an object of the present disclosure to ameliorate one or more problems of the prior art or to at least provide a useful alternative.
An object of the present disclosure is to provide a system for retinal vein occlusion (RVO) detection.
Another object of the present disclosure is to provide a system that seeks to enable faster and more accurate insights from distributed datasets, improving overall analytical capabilities in fields like healthcare, finance, and technology.
Still another object of the present disclosure is to provide a system that facilitates continuous learning from incoming data, ensuring that the analytical models remain up-to-date with the latest trends and patterns.
Yet another object of the present disclosure is to provide a system that reduces the risks associated with data breaches by enabling data analysis without the need to centralize sensitive information, thus maintaining compliance with privacy regulations.
Still another object of the present disclosure is to provide a system that aims to optimize the transmission of information between collaborating entities, reducing the bandwidth requirements and enhancing the speed of the learning process.
Yet another object of the present disclosure is to provide a system that develops scalable, allowing easy integration with multiple data sources across different domains while maintaining efficient performance.
Still another object of the present disclosure is to provide a system that intends to lower the computational burden on individual data sources, making it accessible to institutions with varying levels of technological infrastructure.
Yet another object of the present disclosure is to provide a system that adapts to various sectors beyond healthcare, including areas like financial services, manufacturing, and autonomous systems, where data-driven insights are critical.
Still another object of the present disclosure is to provide a method for retinal vein occlusion (RVO) detection.
Other objects and advantages of the present disclosure will be more apparent from the following description, which is not intended to limit the scope of the present disclosure.
SUMMARY
The present disclosure provides a system and method for retinal vein occlusion (RVO) detection, the system comprising: a plurality of distributed client nodes, and a central aggregation server.
The plurality of distributed client nodes, each node comprising: a local dataset, and a local processing module.
The local dataset of retinal images is categorized into at least two classes representing healthy and RVO-affected eyes.
The local processing module is configured to train a Convolutional Neural Network (CNN) on the dataset.
The central aggregation server comprises a federated learning module and a technique selection module.
The federated learning module is configured to coordinate training across client nodes.
The technique selection module is configured to dynamically select between a Federated Averaging (FedAvg) technique for Independent and Identically Distributed (IID) data and a stochastic controlled averaging technique for federation learning technique for non-IID data.
In an embodiment, the federated learning module further comprises:
o a communication layer configured to facilitate secure data exchange between the client nodes and the aggregation server using encrypted transmission protocols; and
o a differential privacy mechanism configured to prevent leakage of sensitive information during the data exchange process.
In an embodiment, the FedAvg technique is configured to:
o receive gradient updates from each client node based on their local CNN training;
o aggregate these gradient updates by weighted averaging to produce a global model that is less sensitive to biases introduced by individual client datasets; and
o iteratively refine the global model by sending updated parameters back to the client nodes for continuous learning.
In an embodiment, the Federated Learning technique is configured to manage client drift by incorporating control variates, comprising:
o a local control variate update mechanism is configured to reduce the deviation of local model updates from the global model by applying a correction term; and
o a synchronization process at the central server is configured to aggregate the differences in local model updates and apply these adjustments to the global model, enhancing convergence speed in non-IID scenarios.
In an embodiment, the system includes a communication layer is further configured to:
o implement a secure aggregation technique using homomorphic encryption to prevent unauthorized access to model updates; and
o optimize data transmission by compressing the model updates before sending them to the central server, reducing bandwidth usage.
In an embodiment, the system includes data preprocessing, and the augmentation module is configured to:
o implement adaptive data augmentation techniques that dynamically modify augmentation parameters based on the quality of the training dataset; and
o utilize synthetic image generation techniques to enhance data diversity in underrepresented classes, improving CNN's robustness and accuracy.
In an embodiment, the system includes the control variates within the Federated Learning technique that are configured to:
o dynamically adjust the local learning rate based on the magnitude of updates to the global model, ensuring smooth convergence across diverse datasets;
o implement a regularization technique to prevent overfitting of local models to the specific datasets of individual client nodes.
In an embodiment, the system includes a data security module within the federated learning module, which is configured to:
o detect and mitigate potential data poisoning attacks that may be attempted by compromised client nodes; and
o utilize blockchain technology to verify the integrity of model updates before they are aggregated by the central server.
In an embodiment, the system further comprises a robustness enhancement module configured to:
o introduce noise to the training data at each client node to improve the resilience of the CNN against adversarial attacks; and
o implement dropout and other regularization techniques in the CNN architecture to prevent model overfitting.
In an embodiment, the system is further configured to enhance computational efficiency through:
o the use of control variates is configured to reduce redundant computations in the federated learning technique, lowering the overall computational load on client nodes;
o synchronizing local and global model updates asynchronously to optimize resource utilization across the federated learning network.
The present disclosure provides a method for dynamically optimizing federated learning for retinal image classification, further comprising:
o detecting the distribution pattern of data across the client nodes;
o automatically selecting the FedAvg technique configured to leverage uniform data distribution for fast convergence in IID scenarios; and
o switching to the Stochastic Controlled Averaging technique configured to stabilize the training process and mitigate client drift in non-IID data scenarios.
In an embodiment, the method further comprises the following steps:
o tracking, by a real-time monitoring module, the training progress and data distribution patterns of each client node; and
o identifying, by an anomaly detection sub-module, irregular data patterns or malicious activities that may compromise the training process.
In an embodiment, the method includes the real-time monitoring module is further configured to:
o detect client dropout events and dynamically reassign training workloads to ensure uninterrupted progress in model training; and
o implement a fault-tolerance mechanism to prevent data loss or model degradation in case of node failures.
In an embodiment, the method further comprising a model evaluation and feedback loop, is configured to:
o assess the performance of the global model using validation data at the client nodes and provide feedback on model accuracy, precision, and recall metrics; and
o adjust the hyperparameters of the CNN architecture dynamically based on the evaluation results to optimize model performance.
In an embodiment, a computer-implemented process for enhancing retinal vein occlusion (RVO) detection through federated learning, wherein the process comprises:
o training a local CNN model at each client node on retinal image data to distinguish between normal and RVO-affected conditions;
o aggregating local updates at a central server configured to mitigate the influence of outlier nodes using weighted average techniques; and
o utilizing the stochastic controlled averaging technique configured to apply control variates that guide the global model's convergence in non-IID settings, improving diagnostic accuracy.
In an embodiment, a specialized data distribution framework for non-IID scenarios in federated learning, wherein the framework is configured to:
o classify clients into hierarchical tiers based on data quality and relevance, assigning a higher weight to updates from nodes with higher data integrity; and
o adaptively redistribute training tasks among the client nodes based on their computational capabilities, optimizing resource utilization across the network.
In an embodiment, a federated learning-based framework for RVO detection, wherein the framework is configured to support continuous learning by:
o incorporating newly collected retinal image data into local models without retraining from scratch, maintaining the global model's relevance and adaptability to evolving data trends; and
o implementing a model validation and performance feedback loop is configured to adjust hyperparameters based on real-time training metrics.
BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWING
A system and method for retinal vein occlusion (RVO) detection, of the present disclosure will now be described with the help of the accompanying drawing in which:
Figure 1 illustrates a block diagram of the present disclosure;
Figure 2 illustrates a flowchart of the process, in accordance with the present disclosure;
Figure 3 illustrates a retinal vein occlusion detection in an IID scenario using the Federated Average approach, in accordance with the present disclosure;
Figure 4 illustrates a retinal vein occlusion detection in an non-IID scenario using a stochastic controlled averaging technique approach, in accordance with the present disclosure;
Figure 5 illustrates Convolutional Neural Network architecture, in accordance with the present disclosure; and
Figure 6 illustrates a left-side (normal) and right-side(critical) of Retinal vein occlusion (RVO), in accordance with the present disclosure.
LIST OF REFERENCE NUMERALS
100 - System
110 - Distributed client nodes.
112 - Local dataset of retinal images.
114 - Local processing module.
120 - Central aggregation server.
122 - Federated learning module.
124 - Technique selection module.
126 - Federated Averaging FedAvg technique.
128 - Stochastic Controlled Averaging technique
130 - Communication layer.
132 - Differential privacy mechanism.
134 - Global model.
136 - Local control variate update mechanism.
138 - Synchronization process
140 - Data preprocessing and augmentation module.
142 - Real-time monitoring module.
144 - Anomaly detection sub-module.
146 - Data security module.
148 - Model evaluation and performance feedback loop.
150 - Robustness enhancement module.
200 - Method
DETAILED DESCRIPTION
The present disclosure relates to machine learning and artificial intelligence, specifically to methods and systems for enhancing the accuracy of medical image analysis using distributed learning techniques. The disclosure involves techniques that optimize model training across multiple data sources without compromising data privacy, applicable in fields requiring data-intensive pattern recognition.
Embodiments, of the present disclosure, will now be described with reference to the accompanying drawing.
Embodiments are provided so as to thoroughly and fully convey the scope of the present disclosure to the person skilled in the art. Numerous details are set forth, relating to specific components, and methods, to provide a complete understanding of embodiments of the present disclosure. It will be apparent to the person skilled in the art that the details provided in the embodiments should not be construed to limit the scope of the present disclosure. In some embodiments, well known processes, well known apparatus structures, and well known techniques are not described in detail.
The terminology used, in the present disclosure, is only for the purpose of explaining a particular embodiment and such terminology shall not be considered to limit the scope of the present disclosure. As used in the present disclosure, the forms "a," "an," and "the" may be intended to include the plural forms as well, unless the context clearly suggests otherwise. The terms "including," and "having," are open ended transitional phrases and therefore specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not forbid the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. The particular order of steps disclosed in the method and process of the present disclosure is not to be construed as necessarily requiring their performance as described or illustrated. It is also to be understood that additional or alternative steps may be employed.
When an element is referred to as being "engaged to," "connected to," or "coupled to" another element, it may be directly engaged, connected, or coupled to the other element. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed elements.
Referring Figure 1, the present disclosure provides a system (100) for retinal vein occlusion (RVO) detection, said system (100) comprising: a plurality of distributed client nodes (110), and a central aggregation server (120).
The plurality of distributed client nodes (110), each node comprising: a local dataset (112), a local processing module (114).
The local dataset (112) of retinal images is categorized into at least two classes representing healthy and RVO-affected eyes.
The local processing module (114) is configured to train a Convolutional Neural Network (CNN) on the dataset.
The central aggregation server (120) comprises a federated learning module (122) and a technique selection module (124).
The federated learning module (122) is configured to coordinate training across client nodes (110).
The technique selection module (124) is configured to dynamically select between a federated averaging (FedAvg) technique (126) for independent and identically distributed (IID) data and a stochastic controlled averaging technique (128) for non-IID data.
In an embodiment, the federated learning module (122) further comprises:
• a communication layer (130) configured to facilitate secure data exchange between the client nodes (110) and the aggregation server (120) using encrypted transmission protocols; and
• a differential privacy mechanism (132) configured to prevent leakage of sensitive information during the data exchange process.
In an embodiment, the FedAvg technique (126) is configured to:
• receive gradient updates from each client node (110) based on their local CNN training; and
• aggregate these gradient updates by weighted averaging to produce a global model (134) that is less sensitive to biases introduced by individual client datasets;
• iteratively refine the global model (134) by sending updated parameters back to the client nodes (110) for continuous learning.
In an embodiment, the stochastic controlled averaging (128) is configured to manage client drift by incorporating control variates, comprising:
• a local control variate update mechanism (136) configured to reduce the deviation of local model updates from the global model (134) by applying a correction term; and
• a synchronization process (138) at the central server (120) configured to aggregate the differences in local model updates and apply these adjustments to the global model, enhancing convergence speed in non-IID scenarios.
In an embodiment, the system (100) includes a communication layer (130) is further configured to:
• implement a secure aggregation technique using homomorphic encryption to prevent unauthorized access to model updates; and
• optimize data transmission by compressing the model updates before sending them to the central server (120), reducing bandwidth usage.
In an embodiment, the system (100) includes data preprocessing and the augmentation module (140) is configured to:
• implement adaptive data augmentation techniques that dynamically modify augmentation parameters based on the quality of the training dataset; and
• utilize synthetic image generation techniques to enhance data diversity in underrepresented classes, improving CNN's robustness and accuracy.
In an embodiment, the system (100) includes the control variates within the Stochastic Controlled Averaging (128) are configured to:
• dynamically adjust the local learning rate based on the magnitude of updates to the global model (134), ensuring smooth convergence across diverse datasets; and
• implement a regularization technique to prevent overfitting of local models to the specific datasets of individual client nodes (110).
In an embodiment, the system (100) includes a data security module (146) within the federated learning module (122), configured to:
• detect and mitigate potential data poisoning attacks that may be attempted by compromised client nodes (110); and
• utilize blockchain technology to verify the integrity of model updates before they are aggregated by the central server (120).
In an embodiment, the system (100) further comprises a robustness enhancement module (150) configured to:
• introduce noise to the training data at each client node (110) to improve the resilience of the CNN against adversarial attacks; and
• implement dropout and other regularization techniques in the CNN architecture to prevent model overfitting.
In an embodiment, the system (100) is further configured to enhance computational efficiency through:
• The use of control variates (136) configured to reduce redundant computations in the Stochastic Controlled Averaging (128), lowering the overall computational load on client nodes (110);
• Synchronizing local and global model updates asynchronously to optimize resource utilization across the federated learning network.
In an embodiment, a computer-implemented process for enhancing retinal vein occlusion (RVO) detection through federated learning, wherein the process comprises:
• training a local CNN model at each client node (110) on retinal image data to distinguish between normal and RVO-affected conditions;
• aggregating local updates at a central server (120) configured to mitigate the influence of outlier nodes using weighted average techniques; and
• utilizing the stochastic controlled averaging technique (128) configured to apply control variates that guide the global model's convergence in non-IID settings, improving diagnostic accuracy.
In an embodiment, a specialized data distribution framework for non-IID scenarios in federated learning, wherein the framework is configured to:
• classify clients (110) into hierarchical tiers based on data quality and relevance, assigning a higher weight to updates from nodes with higher data integrity; and
• adaptively redistribute training tasks among the client nodes (110) based on their computational capabilities, optimizing resource utilization across the network.
In an embodiment, a federated learning-based framework for RVO detection, wherein the framework is configured to support continuous learning by:
• incorporating newly collected retinal image data into local models (114) without retraining from scratch, maintaining the global model's (134) relevance and adaptability to evolving data trends; and
• implementing a model validation and performance feedback loop (148) configured to adjust hyperparameters based on real-time training metrics.
Figure 2 illustrates a flowchart that includes the steps involved in a method (200) for retinal vein occlusion (RVO) detection, in accordance with an embodiment of the present disclosure. The order in which method (200) is described is not intended to be construed as a limitation, and any number of the described method (200) steps may be combined in any order to implement method (200), or an alternative method. Furthermore, method (200) may be implemented by processing resource or electronic device(s) through any suitable hardware, non-transitory machine-readable medium/instructions, or a combination thereof. The method (200) comprises the following steps:
At step (202), the method (200), includes detecting the distribution pattern of data across the client nodes (110).
At step (204), the method (200), includes automatically selecting the FedAvg technique (126) configured to leverage uniform data distribution for fast convergence in IID scenarios.
At step (206), the method (200), includes switching to the Stochastic Controlled Averaging technique (128) is configured to stabilize the training process and mitigate client drift in non-IID data scenarios.
In an embodiment, the method (200) further comprises the following steps:
• tracking, by a real-time monitoring module (142), the training progress and data distribution patterns of each client node (110); and
• identifying, by an anomaly detection sub-module (144), irregular data patterns or malicious activities that may compromise the training process.
In an embodiment, the method (200) includes the real-time monitoring module (142) is further configured to:
• detect client dropout events and dynamically reassign training workloads to ensure uninterrupted progress in model training; and
• implement a fault-tolerance mechanism to prevent data loss or model degradation in case of node failures.
In an embodiment, the method (200) further comprising a model evaluation and feedback loop (148), is configured to:
• assess the performance of the global model (134) using validation data at the client nodes (110) and provide feedback on model accuracy, precision, and recall metrics; and
• adjust the hyperparameters of the CNN architecture dynamically based on the evaluation results to optimize model performance.
Figure 3 illustrates a retinal vein occlusion detection in an IID scenario using the Federated Average approach, in accordance with the present disclosure; Figure 3 depicts a federated learning setup where multiple hospitals (labelled as Hospital A, Hospital B, and Hospital N) collaboratively train machine learning models without sharing their private data. Each hospital has its local data that undergoes preprocessing and data augmentation before training a local model. The local parameters of each model are then computed and sent to a central server that aggregates these parameters. The aggregation is done using a weighted averaging formula, which combines the contributions from each hospital to create an updated global model. This global model is subsequently sent back to the hospitals for further local training. The process iterates, with each hospital updating its local model based on the newly updated global model. The diagram highlights that private data remains strictly localized at each hospital and is not exchanged, ensuring data privacy throughout the collaborative training process.
Figure 4 illustrates a retinal vein occlusion detection in an non-IID scenario using a stochastic controlled averaging technique approach, in accordance with the present disclosure. Figure 4, it involves multiple hospitals (labelled as Hospital A, Hospital B, and Hospital N) that individually train local models on their data after preprocessing and data augmentation. Each hospital computes differences in local model updates (Δwi) and control variates (Δci) to account for the variability in their data. These differences are then sent to a central server, which aggregates the model update differences and control variate differences to refine the global model. The server also maintains global control variates (C) to stabilize learning across the client nodes. The updated global model and control variates are sent back to each hospital to guide further local training. Throughout the process, private data remains localized at each hospital, ensuring that sensitive information is never shared, which enhances data privacy and security.
Figure 5 illustrates Convolutional Neural Network architecture, in accordance with the present disclosure. Figure 5 describes the architecture consisting of four convolutional layers (Conv 1 to Conv 4) followed by four max pooling layers (Pool 1 to Pool 4), which are designed to extract relevant features from the input retinal image. These layers progressively reduce the spatial dimensions while enhancing important visual features that are crucial for classification. After the convolutional and pooling layers, the network includes a fully connected layer (FC), which integrates the extracted features into a single output. The final output layer provides a binary classification, with "0" indicating a healthy eye and "1" indicating an RVO-affected eye. This sequential structure enables the model to accurately analyse retinal images and distinguish between different eye conditions.
Figure 6 illustrates a left-side (normal) and right-side(critical) of Retinal vein occlusion (RVO), in accordance with the present disclosure describing the Convolutional Neural Network (CNN) architecture used for retinal vein occlusion (RVO) detection. The image (Fig 4) shows a comparison between a normal retina and a retina affected by retinal vein occlusion (RVO). The left side displays a "0-Normal" eye, indicating a healthy retina with no visible abnormalities. The right side illustrates a "1-Critical" eye, where the retina is affected by RVO, characterized by visible blood vessel blockages and irregularities. This visual distinction between the normal and critical conditions is crucial for the system's machine learning model to accurately identify and classify retinal images, helping in the early detection and treatment of RVO.
Table 1: Number of samples in the dataset before and after Augmentation

Dataset Classes Total
Normal Critical
RVO
(Before Augmentation) 2,995 440 3,435
RVO
(After Augmentation) 2,995 2,960 5,955

Embodiment 1: Dataset Preparation and Distribution
Referring to Table 1, In one embodiment, the system for retinal vein occlusion (RVO) detection utilizes a dataset containing retinal images categorized into two classes: normal and RVO-affected (critical). Before data augmentation, the dataset includes 2,995 images of normal eyes and 440 images of critical eyes, totalling 3,435 images. After augmentation, the dataset grows to 2,995 normal images and 2,960 critical images, totalling 5,955 images. The dataset is distributed among five clients in both IID and non-IID manners. In the IID scenario, each client receives an equal and balanced portion of normal and critical images, ensuring a uniform data distribution across all clients. Conversely, in the non-IID scenario, each client receives an imbalanced portion of normal and critical images, to simulate real-world data distribution scenarios.
Embodiment 2: Data Preprocessing
In another embodiment, data preprocessing is applied to all retinal images to ensure uniformity and consistency. Each image is resized to a standard resolution of 224x224 pixels. This resizing step ensures that the input dimensions are consistent across the entire dataset, facilitating stable model performance. Additionally, the pixel values are normalized by scaling them between 0 and 1, which helps accelerate the training process by avoiding numerical instability and slow convergence.
Embodiment 3: Data Augmentation
Table 2: Data Augmentation Parameters and Values
Augmented Parameter Value
Horizontal Flip True
Rotation Range 20
Width-Shift Range 0.2
Height-Shift Range 0.2








Table 2, in this embodiment, focuses on data augmentation techniques designed to address class imbalances within the dataset. Techniques such as horizontal flipping, rotation (up to 20 degrees), and width and height shifting (0.2 units) are employed to artificially expand the dataset. These augmentations increase data diversity, enabling the model to better generalize and improve its accuracy in distinguishing between normal and RVO-affected retinal images.
Embodiment 4: Local Model Training using CNN Architecture
In another embodiment, the local training process is performed using a Convolutional Neural Network (CNN) with a specific architecture. This CNN consists of four convolutional layers activated by the ReLU function, followed by max-pooling layers to reduce the spatial dimensions. Each convolutional layer extracts features from the input image, with the first layer containing 64 filters and the subsequent layers containing 128 filters. The network includes a fully connected layer with 128 units, a dropout layer with a rate of 0.5 to prevent overfitting, and a final SoftMax layer for binary classification between normal and RVO-affected classes.
Embodiment 5: Model Initialization and Training using FedAvg Technique
In this embodiment, the system utilizes the Federated Averaging (FedAvg) technique to train the model when data is Independent and Identically Distributed (IID). Each client initializes its local model parameters and calculates the gradients of the loss function with respect to these parameters. The clients then update their model parameters using gradient descent and share these updated parameters with the central server. The server aggregates the updates from all clients using a weighted average to form the global model, which is then redistributed to the clients for further training.
Embodiment 6: Model Training using a stochastic controlled averaging technique
In another embodiment, when the data is non-IID, the system employs the stochastic controlled averaging technique to manage client drift and stabilize the training process. Each client computes local gradients and updates its model parameters using the following formula:
wi=wi− ηl (∂L/∂wi − ci + C)
where:
wi represents the parameters of the local model for client i.
ηl is the local learning rate.
is the gradient of the loss function with respect to the model parameters.
is the control variate for client i.
is the global control variate.
Embodiment 7: Control Variate Update and Difference Calculation
This embodiment describes how each client updates its control variate during the STOCHASTIC CONTROLLED AVERAGING TECHNIQUE training process using the following equation:

where:
ci+ represents the updated control variate for client i.
k is the number of local epochs.
X is the global model parameters.
wi is the local model parameters for client i.
The client then calculates the differences between the updated local model parameters and the global model parameters as follows:

Embodiment 8: Server-Side Aggregation and Global Model Update
In another embodiment, the central server aggregates the differences in local model parameters and control variates received from all clients. The server performs the following steps:
Global Gradient Aggregation:

where g is the average global gradient based on the local parameter differences Δwi.
Global Control Variate Aggregation:

where ΔC is the average difference in control variates.

Global Model Update:

where ηg is the global learning rate.
Global Control Variate Update: The global control variate C is updated based on the weighted average of the control variate differences ΔC.
The server then communicates the updated global model parameters X and global control variate C back to the clients for the next iteration of training.
Embodiment 7: Control Variate Update and Difference Calculation
This embodiment describes how each client updates its control variate during the stochastic controlled averaging technique training process. The control variance for each client is adjusted using the difference between the global model and the local model parameters. This approach helps synchronize local model updates with the global model, minimizing deviations and ensuring that the training remains stable across multiple rounds.
Embodiment 8: Server-Side Aggregation and Global Model Update
In another embodiment, the central server aggregates the differences in local model parameters and control variates received from all clients. It computes the average difference in local model parameters and updates the global model accordingly. The server also updates the global control variate to maintain synchronization across all client nodes. The updated global model and control variate are then sent back to the clients for the next training iteration.
Embodiment 9: Adaptive Data Handling for Client Nodes
This embodiment focuses on how the system dynamically adjusts to varying data distributions across client nodes. It highlights the system's ability to adapt the learning process based on the capabilities of each client node, optimizing resource utilization. Clients with imbalanced data distributions are handled more effectively using the control variate mechanism, ensuring that their contributions to the global model are accurate and consistent.
Embodiment 10: Privacy Preservation and Communication Efficiency
In another embodiment, the system emphasizes privacy-preserving mechanisms that ensure sensitive data from client nodes is never shared directly with the central server. Techniques such as differential privacy and secure aggregation are utilized to protect data integrity. The system also minimizes communication overhead by transmitting only the differential updates of model parameters, significantly reducing the amount of data exchanged between the server and clients.
Embodiment 11: Continuous Model Improvement and Convergence
This embodiment explains how the system supports continuous model training and convergence. By iteratively updating both the local and global models using either the FedAvg (126) or stochastic controlled averaging technique (128), the system ensures that the global model improves over time. This iterative process continues until the model reaches an optimal level of accuracy in detecting retinal vein occlusion.
In the operative configuration, the system (100) for retinal vein occlusion (RVO) detection operates as a distributed network with multiple client nodes (110) and a central aggregation server (120). Each client node (110) contains a local dataset (112) of retinal images and a processing module (114) that trains a convolutional neural network (CNN) on its data. The federated learning module (122) on the server coordinates this training by dynamically selecting either the federated averaging (FedAvg) technique (126) for IID data or the stochastic controlled averaging technique (128) for non-IID data, based on data distribution patterns. The communication layer (130) securely facilitates the exchange of model updates without transmitting raw data, while the synchronization process (138) aggregates these updates and applies control variates to the global model (134), enhancing convergence speed and model accuracy.
Advantageously, the system (100) leverages federated learning to protect data privacy by ensuring that sensitive retinal image data remains within each client node (110), reducing the risks of data breaches. The stochastic controlled averaging technique (128) effectively addresses the challenge of client drift in non-IID data scenarios, promoting faster convergence and stable training across diverse datasets. By incorporating control variates, the system dynamically adjusts learning rates to prevent overfitting and improve model stability. The data preprocessing and augmentation module (140) further strengthens CNN's ability to handle variations in image quality, while real-time monitoring and fault-tolerance mechanisms enhance system reliability. These features collectively make the system robust, efficient, and highly adaptable for accurate RVO detection across different medical institutions.
The functions described herein may be implemented in hardware, executed by a processor, firmware, or any combination thereof. Other examples and implementations are within the scope and spirit of the disclosure and appended claims. The nature of the disclosure, it can be implemented by a processor, hardware, firmware, hardwiring, or combinations of any of these. Features implementing functions may also be physically located at various positions, including being distributed such that portions of functions are implemented at different physical locations.
The foregoing description of the embodiments has been provided for purposes of illustration and is not intended to limit the scope of the present disclosure. Individual components of a particular embodiment are generally not limited to that particular embodiment, but, are interchangeable. Such variations are not to be regarded as a departure from the present disclosure, and all such modifications are considered to be within the scope of the present disclosure.
TECHNICAL ADVANCEMENTS
The present disclosure described hereinabove has several technical advantages including, but not limited to, a system and method for retinal vein occlusion (RVO) detection, which;
• seeks to enable faster and more accurate insights from distributed datasets, improving overall analytical capabilities in fields like healthcare, finance, and technology.
• facilitates continuous learning from incoming data, ensuring that the analytical models remain up-to-date with the latest trends and patterns;
• reduces the risks associated with data breaches by enabling data analysis without the need to centralize sensitive information, thus maintaining compliance with privacy regulations;
• aims to optimize the transmission of information between collaborating entities, reducing the bandwidth requirements and enhancing the speed of the learning process;
• develops scalable, allowing easy integration with multiple data sources across different domains while maintaining efficient performance.
• intends to lower the computational burden on individual data sources, making it accessible to institutions with varying levels of technological infrastructure; and
• adapts to various sectors beyond healthcare, including areas like financial services, manufacturing, and autonomous systems, where data-driven insights are critical.
The foregoing disclosure has been described with reference to the accompanying embodiments which do not limit the scope and ambit of the disclosure. The description provided is purely by way of example and illustration.
The embodiments herein and the various features and advantageous details thereof are explained with reference to the non-limiting embodiments in the following description. Descriptions of well known components and processing techniques are omitted so as to not unnecessarily obscure the embodiments herein. The examples used herein are intended merely to facilitate an understanding of ways in which the embodiments herein may be practiced and to further enable those of skill in the art to practice the embodiments herein. Accordingly, the examples should not be construed as limiting the scope of the embodiments herein.
The foregoing description of the specific embodiments so fully reveal the general nature of the embodiments herein that others can, by applying current knowledge, readily modify and/or adapt for various applications such specific embodiments without departing from the generic concept, and, therefore, such adaptations and modifications should and are intended to be comprehended within the meaning and range of equivalents of the disclosed embodiments. It is to be understood that the phraseology or terminology employed herein is for the purpose of description and not of limitation. Therefore, while the embodiments herein have been described in terms of preferred embodiments, those skilled in the art will recognize that the embodiments herein can be practiced with modification within the spirit and scope of the embodiments as described herein.
Any discussion of devices, articles or the like that has been included in this specification is solely for the purpose of providing a context for the disclosure. It is not to be taken as an admission that any or all of these matters form a part of the prior art base or were common general knowledge in the field relevant to the disclosure as it existed anywhere before the priority date of this application.
While considerable emphasis has been placed herein on the components and component parts of the preferred embodiments, it will be appreciated that many embodiments can be made and that many changes can be made in the preferred embodiments without departing from the principles of the disclosure. These and other changes in the preferred embodiment as well as other embodiments of the disclosure will be apparent to those skilled in the art from the disclosure herein, whereby it is to be distinctly understood that the foregoing descriptive matter is to be interpreted merely as illustrative of the disclosure and not as a limitation. , Claims:WE CLAIM:
1. A system (100) for retinal vein occlusion (RVO) detection, said system (100) comprising:
• a plurality of distributed client nodes (110), each node comprising:
o a local dataset (112) of retinal images categorized into at least two classes representing healthy and RVO-affected eyes; and
o a local processing module (114) configured to train a convolutional neural network (CNN) on the dataset;
• a central aggregation server (120) comprising:
o a federated learning module (122) configured to coordinate training across client nodes (110); and
o a technique selection module (124) configured to dynamically select between a Federated Averaging (FedAvg) technique (126) for independent and identically distributed (IID) data and a stochastic controlled averaging technique for Stochastic Controlled Averaging technique (128) for non-IID data.
2. The system (100) of claim 1, wherein the federated learning module (122) further comprises:
o a communication layer (130) configured to facilitate secure data exchange between the client nodes (110) and the aggregation server (120) using encrypted transmission protocols; and
o a differential privacy mechanism (132) configured to prevent leakage of sensitive information during the data exchange process.
3. The system (100) of claim 1, wherein the FedAvg technique (126) is configured to:
o receive gradient updates from each client node (110) based on their local CNN training; and
o aggregate these gradient updates by weighted averaging to produce a global model (134) that is less sensitive to biases introduced by individual client datasets;
o iteratively refine the global model (134) by sending updated parameters back to the client nodes (110) for continuous learning.
4. The system (100) of claim 1, wherein the stochastic controlled averaging technique (128) is configured to manage client drift by incorporating control variates, comprising:
o a local control variate update mechanism (136) configured to reduce the deviation of local model updates from the global model (134) by applying a correction term; and
o a synchronization process (138) at the central server (120) configured to aggregate the differences in local model updates and apply these adjustments to the global model, enhancing convergence speed in non-IID scenarios.
5. The system (100) of claim 1, wherein the communication layer (130) is further configured to:
o implement a secure aggregation technique using homomorphic encryption to prevent unauthorized access to model updates; and
o optimize data transmission by compressing the model updates before sending them to the central server (120), reducing bandwidth usage.
6. The system (100) of claim 1, wherein the data preprocessing and augmentation module (140) is configured to:
o implement adaptive data augmentation techniques that dynamically modify augmentation parameters based on the quality of the training dataset; and
o utilize synthetic image generation techniques to enhance data diversity in underrepresented classes, improving CNN's robustness and accuracy.
7. The system (100) of claim 1, wherein the control variates within the stochastic controlled averaging technique (128) is configured to:
o dynamically adjust the local learning rate based on the magnitude of updates to the global model (134), ensuring smooth convergence across diverse datasets;
o implement a regularization technique to prevent overfitting of local models to the specific datasets of individual client nodes (110).
8. The system (100) of claim 1, wherein the system (100) includes a data security module (146) within the federated learning module (122), configured to:
o detect and mitigate potential data poisoning attacks that may be attempted by compromised client nodes (110); and
o utilize blockchain technology to verify the integrity of model updates before they are aggregated by the central server (120).
9. The system (100) of claim 1, further comprising a robustness enhancement module (150) configured to:
o introduce noise to the training data at each client node (110) to improve the resilience of the CNN against adversarial attacks; and
o implement dropout and other regularization techniques in the CNN architecture to prevent model overfitting.
10. The system (100) of claim 1, further configured to enhance computational efficiency through:
o the use of control variates (136) configured to reduce redundant computations in the stochastic controlled averaging technique (128), lowering the overall computational load on client nodes (110);
o synchronizing local and global model updates asynchronously to optimize resource utilization across the federated learning network.
11. A method (200) for dynamically optimizing federated learning for retinal image classification, further comprising:
o detecting the distribution pattern of data across the client nodes (110);
o automatically selecting the FedAvg technique (126) configured to leverage uniform data distribution for fast convergence in IID scenarios; and
o switching to the stochastic controlled averaging technique (128) configured to stabilize the training process and mitigate client drift in non-IID data scenarios.
12. The method (200) of claim 11, further comprises the following steps:
o tracking, by a real-time monitoring module (142), the training progress and data distribution patterns of each client node (110); and
o identifying, by an anomaly detection sub-module (144), irregular data patterns or malicious activities that may compromise the training process.
13. The method (200) of claim 11, wherein the real-time monitoring module (142) is further configured to:
o detect client dropout events and dynamically reassign training workloads to ensure uninterrupted progress in model training; and
o implement a fault-tolerance mechanism to prevent data loss or model degradation in case of node failures.
14. The method (200) of claim 11, further comprising a model evaluation and feedback loop (148), is configured to:
o assess the performance of the global model (134) using validation data at the client nodes (110) and provide feedback on model accuracy, precision, and recall metrics; and
o adjust the hyperparameters of the CNN architecture dynamically based on the evaluation results to optimize model performance.
15. A computer-implemented process for enhancing retinal vein occlusion (RVO) detection through federated learning, wherein the process comprises:
o training a local CNN model at each client node (110) on retinal image data to distinguish between normal and RVO-affected conditions;
o aggregating local updates at a central server (120) configured to mitigate the influence of outlier nodes using weighted average techniques; and
o utilizing the stochastic controlled averaging technique (128) configured to apply control variates that guide the global model's convergence in non-IID settings, improving diagnostic accuracy.
16. A specialized data distribution framework for non-IID scenarios in federated learning, wherein the framework is configured to:
o classify clients (110) into hierarchical tiers based on data quality and relevance, assigning a higher weight to updates from nodes with higher data integrity; and
o adaptively redistribute training tasks among the client nodes (110) based on their computational capabilities, optimizing resource utilization across the network.
17. A federated learning-based framework for RVO detection, wherein the framework is configured to support continuous learning by:
o incorporating newly collected retinal image data into local models (114) without retraining from scratch, maintaining the global model's (134) relevance and adaptability to evolving data trends; and
o implementing a model validation and performance feedback loop (148) configured to adjust hyperparameters based on real-time training metrics.
Dated this 29th Day of October, 2024

_______________________________
MOHAN RAJKUMAR DEWAN, IN/PA - 25
OF R. K. DEWAN & CO.
AUTHORIZED AGENT OF APPLICANT

TO,
THE CONTROLLER OF PATENTS
THE PATENT OFFICE, AT CHENNAI

Documents