Efficient Federated Aggregation Technique for Brain Tumor Classification Using Transfer Learning Approaches

Federated learning (FL) is a decentralized machine learning paradigm that enables machine learning models to be trained collaboratively across multiple nodes while preserving data privacy. This approach allows the organizations to share and contribute to training the ML models without sharing the data [1]. The aggregation methodology adopted in the collaborative approach plays a vital role in the model's performance and reliability. Initially, a simple averaging method was adopted for aggregation. As per the studies and survey, there is a limited contribution in addressing communication efficiency, scalability issue and selecting the appropriate clients who can take part in FL [2, 3].

Significant progress has been made in classifying brain tumors from MRI scans using a variety of machine learning and deep learning techniques. These techniques help doctors plan treatments and improve the accuracy of diagnoses. Gliomas, meningiomas, and pituitary tumors are among the various forms of brain tumors that can be successfully classified by combining convolutional neural networks (CNNs) with other machine learning models [4, 5].

We are trying to adopt a FL approach for brain tumor classification. Since early detection plays a vital role in brain tumor treatment [5, 6]. Hence, a novel approach PC_FedAvg (Priority Based Federated Averaging) is adopted to select a genuine and limited number of clients to take part in FL, which reduces communication rounds between clients and servers and increases the performance of the global model.

CNN is one of the appropriate machine learning models to accurately detect the tumors in the brain for the early start of the treatment [6]. Due to privacy concerns of medical health data and HIPAA guidelines, usually medical institutes are very rigid in sharing the data. Hence federated machine learning approach is suitable for medical health applications. Various aggregation techniques have been proposed such as parameter-based, output-based. In parameter-based aggregation weight parameters are objects for consideration. Few researchers worked on a precision-based FL approach [7-11]. We considered a parameter-based approach for our study. Some studies focused on aggregation with fine tuning and designed new aggregation algorithms based on the application requirements [12-15]. Emerging FL systems now incorporate secure global aggregation techniques designed for heterogeneous device networks, addressing critical security challenges in multi-device environments [16-18].

Medical image classification problems require large amounts of data on which models can be trained, because of the restrictions in sharing data of the patients, transfer learning is the most suitable approach for data-hungry applications. FL aggregation methodologies have been advanced to improve brain tumor classification accuracy while maintaining data privacy across distributed systems [19-21]. We adopted the Federated Transfer learning approach and used Xception, DenseNet201, and VGG19 models to classify brain tumor types.

We propose PC_FedAvg (Priority Based Client Federated Averaging), a novel aggregation algorithm which selects a limited number of clients based on their local model performance results, to take part in federated aggregation.

We adopted Cross-Silo federated Transfer learning approach for brain tumor classification and used the combination of Figshare, SARTAJ and Br35H benchmark datasets to classify the brain tumors into four types: glioma, meningioma, no tumor and pituitary. We analyzed the performance of our proposed PC_FedAvg with widely used aggregation algorithms FedAvg, FedProx and FedAdam. Results shows that PC_FedAvg is communication efficient as it reduces the communication rounds between client and server. We adopted Federated Transfer Learning approach by choosing Pretrained models Xception and VGG19 and DenseNet201.Initially, we compared the performance of VGG19, DenseNet201 and Xception models for brain tumor classification in centralized machine learning approach. Classification results show that the Xception model outperforms VGG19 and DenseNet with Precision 98.4, Recall 98.4, F1 score 98 and accuracy 99.4%, hence the Exception model is adopted for a federated approach for classification.

1.1 FL in healthcare

FL is essentially machine learning in a decentralized form. Conventional machine learning methods often include training models with data that is collected from several edge devices, such as laptops, and smartphones, and then transferred to a central server. This centralized data storage is the learning platform, where machine learning algorithms, such as neural networks, train themselves on aggregated data before making predictions on fresh data [1]. Improved data privacy, lower bandwidth requirements than transmitting raw data, and the capacity to use insights from various data sources without directly accessing the raw data are just a few benefits of this decentralized machine-learning technique [2, 3]. The FL approach includes 4 steps as shown in the general architecture in Figure 1. Client nodes or edge devices has their own dataset available with them and have a global model initially sent by the global server to all nodes.

1.jpg

Figure 1. FL general architecture

Step1: Individual clients send their local parameters(weights) obtained by training the model on local dataset to the Global server.

Step2: The global server updates the global model by incorporating the global weight obtained by taking the average of all weights received (FedAvg).

Step3: Global server sends the global model to all the clients.

Step4: Clients update the local mode.

FL can be classified Based on devices involved and based on the data or models used as shown in Figure 2.

2.jpg

Figure 2. FL classification types

I Based on the devices involved

i) Cross-Silo: Suitable when large number of devices (Especially Homogeneous).

E.g., when different medical institutes taking part in FL process, referred to as cross-silo federation.

ii) Cross-Device: Limited number of devices are involved (Especially Heterogeneous).

E.g., when different wearable devices from different patients are collaborating for model training, referred to as cross-device.

II Based on the data or models used

i) Horizontal: Same features and same sample space.

ii) Vertical: Different features in same sample space.

iii) Federated Transfer Learning: Use of pre-trained machine learning models.

There are many application areas where FL is beneficial such as smart security systems, transportation systems health care, etc. some of the applications are explained below:

Healthcare: Hospitals can collaboratively train models for disease prediction without sharing patient data [1, 2].

Smart Transportation: FL can enhance Smart Transportation systems by aggregating insights from multiple devices while maintaining data confidentiality [1, 2].

1.2 Aggregation techniques used in FL and its limitations

As shown in Figure 1, the central server performs model aggregation on the weights and builds the global model. Aggregation Techniques can be classified into two types [7, 8].

A) Synchronous aggregation: All clients or nodes may not take part in the aggregation process. One can set the limit such as certain number of epochs or time duration to take part in aggregation. Delayed updates are not considered by the server.

B) Asynchronous aggregation: Usually all Clients or nodes take part in the aggregation process, and delay in the model updates can be considered by the central server [7].

Traditional FedAvg algorithm which is widely used adopts a synchronous approach for communication rounds. The complete aggregation process can be represented with help of steps shown below.

Different Aggregation techniques or algorithms used in FL most widely used are discussed here:

FedAvg: FedAvg is the most commonly used weighted averaging technique based on the available data on the client nodes. It is the generalization of FedSGD (Federated Stochastic Gradient decent). More than one local gradient decent update is performed in FedAvg [1, 2, 7, 8]

Initialize the Global Model: Let the global model at round t be represented as: $w^t$ where $w^t$ is the parameter vector of the global model. Start with an initial global model $w^t$ on the central server.

Local Training: Each client k receives the global model $w^t$, performs local training on the dataset and performs local training on its dataset Dk. where $k \in\{1,2, \ldots, K\}$. Each client minimizes a locally regularized objective function, The local optimization (e.g., SGD) updates the model parameters:

$W_k^{t+1}=w^t-\eta \nabla L_k\left(w^t\right)$     (1)

where, $\eta$ is the learning rate and $\nabla L_k\left(w^t\right)$ is the gradient of the loss function $L_k$ with respect to $w^t$ using the local data Dk.

Model Aggregation: After local training, instead of each client sending its updated weights $w_k^{t+1}$ back to the central server in FedAvg, a selected number of clients will update the weights to the server. At each round, FedAvg randomly selects the clients to participate in the aggregation process. Weights are assigned to the clients based on the quantity of data volume available to them. It adopts the synchronous approach. The limitation of FedAvg is it drops the clients if a certain number of epochs are not performed by them [7, 8, 11, 12, 14].

FedProx: FedProx addresses the heterogeneity of devices and performs well when devices with different computational powers are involved the FL. It allows partial work instead of eliminating clients like FedAvg. In FedProx the proximal term is added to the local objective function of clients to make the local updates stabilize. Each client minimizes a locally regularized objective function, adding a proximal term to reduce the discrepancy between its local model and the global model:

$min \left[F_k\left(w_k\right)+\frac{\mu}{2}\left\|w_k-w^t\right\|\right] 2$     (2)

where, $F_k\left(w_k\right)$ is the local objective function for client $k$ (e.g., empirical risk or loss function on $D_k$).

This proximal term $\frac{\mu}{2}\left\|w_k-w^t\right\| 2$ acts as a penalty that limits how much each client’s local model can deviate from the global model.

Clients perform this optimization for a set number of epochs or until convergence.

FedAdam: Adam optimizer is adopted for FL by FedAdam. Parameters are updated using both first and second moment of gradients. The formula for updating is:

$W_{t+1}=W_t-\frac{\eta}{\sqrt{v_t}+\varepsilon} m_t$     (3)

where, $m_t$ is mean of gradients, $v_t$ is variance of the gradients, $e$ is small constant added and $\eta$ is the learning rate [14].

FedAdam address the challenges of data heterogeneity.

FedYogi: This algorithm is based on Yogi optimizer and it is gradient based optimization which address the limitation of FedAdam (excessively large variance) [14, 15]:

$w_{t+1}^k=w_t-\frac{\eta}{\sqrt{v_t^k}+\varepsilon} m_t^k$    (4)

FedNova: FedNova addresses the non-IID distribution of data in the federated set up and enhances the model aggregation phase. This algorithm introduces a technique to normalize the updates from each client on its local iteration number, before updating the global model.

One of the concerns of these algorithms is limited focus on the number of clients taking part in the aggregation process. It is observed that when the number of clients taking part in the aggregation increases, the performance of few (FedAvg) decreases and some require more communication rounds for convergence [1, 14].

Below are the limitations of aggregation techniques used in FL:

Participating Nodes Availability: All clients may not be available during each training round; hence aggregation results may be biased based on available clients

Data Heterogeneity: Clinets usually having non-IID Data distribution with them and traditional algorithms, e.g., FedAvg performs poorly in such cases [1, 2].

Fairness Issues: Clients with smaller datasets may have less influence on the global model and hence global model may be biased.

Communication Overhead: Frequent communication between client and server consumes network bandwidth and time for global model convergence.

Research Significance:

• To identify and emphasize the latest research trends in aggregation techniques used by federated machine learning.
• Design communication efficient aggregation technique for brain tumor classification in 4 types and analyze the performance of the proposed technique with existing approaches.
• Improve the performance of the global model used to classify brain tumors.

We are trying to balance the limitations of FedAvg which struggles with non-IID data and FedProx, which struggles with slow convergence by selecting a limited number of clients based on local model performance. To address the performance limitation of the machine learning model due to appropriate client selection for aggregation, we proposed a novel approach for aggregation called PC_Fed average which selects the limited number of clients based on their overall performance during local training. Detail algorithm is explained in Section 4. Table 1 gives the details of the aggregation algorithm, its limitations and its applications.

We provide extensive evaluations of the proposed PC_FedAvg algorithm using benchmark image classification datasets, demonstrating its robustness to unbalanced data distributions; and also compared the proposed method to Federated Averaging on empirical experiments, and with fewer communication rounds we obtain comparable accuracy.

Table 1. Survey on transfer learning approach for brain tumor classification

By considering the limitations found in the literature survey, we identified that aggregation of local models trained on divergent data can result in poor global model performance across the entire population traditional aggregation algorithms treats all the clients equally while some clients may have better quality or more representative data.

We designed PC_FedAvg (Priority-based Federated Averaging) by selecting a limited number of clients prioritizing those who are likely to benefit in each round of communication between client and server. Which helps in better representation and faster convergence and reduced gradient conflicts.

Some common priority Matrices can be Data Diversity, Model contribution or local accuracy and resource availability or reliability have chosen local accuracy as priority mastics and chosen the 60% clients with highest local model accuracy.

Prioritizing the client selection in FL aids in the following manner non-IID data makes local models go in opposite directions when optimizing. By choosing more aligned or complementary customers, the system eliminates conflicting updates and converges better.

We used pre-trained deep learning models, Xception, DenseNet201, and VGG19 to classify the brain tumors in to four types glioma, meningioma, and no tumor and pituitary. Cross-silo Federated approach is used and hence number of clients taking part in FL is limited to max 5.

3.1 Mathematical representation of PC_FedAvg algorithm

Let

K: Total number of clients.

$D_k$: Local dataset for client k.

$w^t$: Global model weights at communication round t.

$w_k^t$: Client k’s local model weights at round t.

The detailed mathematical representation of PC_FedAvg is explained in Table 3.

Table 3. Mathematical representation of PC_FedAvg algorithm

3.1.1 Details of representation of the selected clients set in PC_FedAvg

Let: K represent set of all available clients K={1,2,3,4,5….K}, St K represent the subset of clients selected for participation in round t, St={k1,k2,k3,….Kn}.

The server aggregates the weights using a weighted average based on each client’s data size

FedAvg Aggregation:

$w^{t+1}=\sum_{k=1}^K \frac{n_k}{n} \omega_k^{t+1}$     (5)

3.1.2 PC_FedAvg aggregation incorporating selected clients

In FedAvg, if only the updates from the selected clients St used to update the global model. The aggregation formula then becomes:

$w^{t+1}=\sum_{k \in s^t}^K: \frac{n_k}{\sum_{j \varepsilon s^t} n_j} \omega_k^{t+1}$    (6)

where, S^t is set of top 60 % of available clients with highest accuracy in that communication round.