Federated Learning for Multi-Center Medical Image Classification Using Deep Learning Models

Physicians employ medical images, such as computed tomography (CT), magnetic resonance imaging (MRI), and X-rays, to create reliable and accurate DL models that help in the diagnosis, detection, and treatment of many diseases. Different metrics are included in these images to aid in the extraction of patient information from the medical images during analysis. However, the objective of the analysis phase is to identify the processed images using a variety of techniques, including machine learning (ML), which are frequently employed in the analysis and feature extraction of medical images used for disease classification and diagnosis [1, 2].

The classification of a multiclass medical image gallbladder disease dataset is used in this current study.  The gallbladder, a vital liver organ, is crucial for digestion but can be affected by diseases like gallstones, cholecystitis, cancer, and others. Early diagnosis is essential for effective treatment, especially in severe cases like gallbladder cancer, which improves survival rates [3].

Deep Learning for medical image interpretation is crucial for improving diagnostic abilities. However, the dispersion of clinical data across healthcare facilities poses a challenge, as it cannot be combined for centralized model training due to privacy laws and data exchange difficulties. This limits the generalizability and effectiveness of diagnostic models for gallbladder issues [4].

The dataset was distributed to multiple clients using two methods: IID (Independent and Identically Distributed) and Non-IID (Independent and Identically Distributed) to replicate a real-world federated learning environment. IID was achieved by randomly dividing the dataset into equal portions for each client, ensuring identical assignments. The most optimal configuration was used to test the relative performance of the FL model [5].

In this work, we used the dataset of gallbladder diseases as the case study to estimate the proposed system model in the field of medical image classification. The dataset used is IID (Independent and Identically Distributed), which is distributed among 3 healthcare centers and is ideal for deep learning algorithms, as it allows for faster convergence and higher accuracy.

Convolutional Neural Networks (CNNs) are a prominent deep learning technique that automatically extracts features from data and improves diagnostic accuracy in medical image analysis tasks. Combining CNN architectures with transfer learning techniques improves image classification performance, recognizing important visual patterns with less preprocessing on raw data. VGG16Net, the largest CNN architecture, has a high computational load [2, 6], which is employed to classify the gallbladder diseases dataset before using FL.

Deep learning methods' predictive abilities are influenced by the amount of training data, which should come from multiple sources [7]. Obtaining sufficient data in the field of medical imaging is a significant difficulty. This difficulty might be solved by collaboration among various institutions. Sharing medical data in a single location raises a number of legal, privacy, technological, and data-ownership issues.

To solve these issues, Federated Learning (FL) enables individual hospitals to benefit from the large datasets of numerous non-affiliated institutions without centralizing the data in one location [8]. FL's privacy-preserving feature allows collaborations across various medical institutes [9] without exchanging medical data between healthcare centers. This helps to preserve privacy and gain model generality through the global model.

In an ideal FL system, medical institutions collaborate using a centralized orchestration cloud server in a trusted execution environment. FL allows hospitals and other medical institutions to ‏keep‏ local data by training a model across multiple medical data centers. The central server maintains a global shared model, co-owned by all participating institutions, while each institution maintains its local version. The server verifies local model quality before aggregating updates based on preset criteria. Global models iterate between aggregation servers and participating institutions, producing high-quality, converged, and performant FL models [10], which are applied to an estimated gallbladder dataset.

Aggregation algorithms are techniques for combining the outcomes of several models that have been trained on the client's end using local data. In addition to updating the global model, they manage the fusion of local client training outcomes. In a FL scenario, several aggregation techniques are employed based on objectives such as convergence rate and user privacy. Weighted aggregation, secure aggregation, momentum aggregation, clipped average aggregation, and average aggregation are a few of these strategies. Average aggregation, which averages client updates, provides benefits like simplicity and increased model accuracy, but it can't work well with non-IID data and is vulnerable to fraudulent clients and outliers. The Federated Averaging FedAvg, FedProx, FedDist, and HeteroSAg aggregation algorithms are being developed for federated learning [11].

The Federated Averaging (FedAvg) is a fundamental algorithm that uses a central server to acquire a primary model, which is then trained using local data from each client. The model parameters are then transmitted to the central server, which then aggregates global models using parameters from other clients. The aggregated global model is then distributed to the clients, marking the end of one learning round. The process continues until the model's accuracy meets the required level, with both clients and the central server maintaining the models from previous iterations [12].

Alternatively, to enable distributed and privacy-preserving FL model training, Local data is used to train the model, and only the model updates are transmitted to a central server. By reducing the amount of data sent between devices and servers and decentralizing the machine learning process, federated learning was able to reduce the risk of data breaches due to malicious attacks and malfunctions in the systems.

Privacy preservation in FL is a crucial aspect of different learning models, ensuring the privacy of training data while enabling successful model training. Robust aggregation techniques, such as FedAvg and secure aggregation, are used to address model integrity attacks, while federal transfer learning fine-tunes models on decentralized data, outlier identification and removal improve performance and model resilience. Different approaches to secure privacy-preserving methods are explored, ranging from simple averaging to more advanced methods such as secure multi-party computing and differential privacy [11, 13].

In this work, a federated deep learning models based on CNN and VGG16 has been proposed for multiple clients (multiple healthcare centers), each client has part of the dataset to train. The gallbladder dataset has been used with eight classes divided between five clients. The rest of the paper is organized as follows: in Section 2 the related work is described. Sections 3 and 4 show a description of the dataset. algorithms and methodologies used in this work are presented in Section 5, while the proposed system design is shown in Section 6. In Section 7, a discussion of the obtained results is presented. Finally, a conclusion is introduced in Section 8.

In the proposed system, The Gallbladder (GB) diseases dataset is a database of high-quality ultrasound pictures taken at four hospitals in Baghdad, Iraq.  The data was gathered over four years from Jenin Hospital, the Al-Numan Teaching Hospital Specialized Gastroenterology Center, and the Gastroenterology Department of the City of Medicine Teaching Hospital. The information is essential for creating deep learning and machine learning algorithms that can identify and categorize gallbladder illnesses. The information supports comparison research and testing of new methods for an ultrasound image analysis to investigate the medical field and enhance patient care. The dataset was collected 10,692 high-resolution ultrasound images of the gallbladder from 1,782 individuals [27]. The images are organized into nine classes, each representing a specific gallbladder disease based on anatomical landmarks. The dataset includes images from female patients (6,246 images, average age 47 years) and male patients (4,446 images, average age 53 years) [28]. The images were acquired using cutting-edge technologies from four different ultrasound machines (Siemens Acuson X700, Philips Affiniti 70, Philips CX50 and Canon Viamo c100). The data collection took place over four years at four medical facilities in Baghdad, Iraq, with medical staff members and expert doctors contributing to the collection [29].

In this current work we are used only eight classes of gallbladder diseases dataset which are used to training and evaluation model, these selected classes or diseases (gallstones, cholecystitis, gangrenous cholecystitis, gallbladder perforation, polyps and cholesterol crystals, gallbladder adenomyomatosis, cancer, and intraabdominal and retroperitoneum problems) of 9702 images each resized to the dimension 224×224 pixels with RGB color mode. We show examples of each class in Figure 1, highlighting the differences in appearance on ultrasound at different gallbladder diseases. Table 1 indicates the whole distribution of Gallbladder disease images in diverse classes within the dataset.

image003.png

Figure 1. Sample ultrasound images from different classes

Table 1. Number of Gallbladder disease images per class

The GB, a hollow organ in the abdomen, stores bile secreted by the liver, causing numerous pathologies, with cholelithiasis being the most common. The list of frequent GB pathologies is summarized below.

Gallstones, which can be mild or severe, are tiny calcium crystals and cholesterol/bile salts that develop inside the gallbladder (GB). Cholecystitis occurs when stones in the bile duct obstruct the bile duct, trapping bile and causing inflammation. Gangrenous cholecystitis is a severe form of cholecystitis that necessitates an immediate cholecystectomy since it causes necrosis of the GB owing to problems with the blood supply. Perforation of the GB wall, a common major complication in inflamed GB, is a serious issue in people with diabetes and compromised immune systems, requiring urgent intervention for treatment. Polyps and cholesterol crystals are common causes of GB disease, with polyps occurring in 2.6%-9.9% of cases. Cholesterol polyps are benign, while endothelial polyps may develop into cancer. A major risk factor for adenomyomatosis of the GB is a persistent infection. The disease is characterized by mucosal epithelial hypertrophy, which leads to Luschka's crypts. Carcinoma is a GB cancer, a rare, female-dominated tumour, is more common in over 70-year-olds, often accompanied by gallstones. Early diagnosis and treatment are crucial for survival. Intraabdominal and retroperitoneum problems, including GB cancer spreading to retroperitoneal structures, can be detected using MRI or CT and ultrasound for further evaluation [29].

The methodology is structured to leverage advanced deep learning techniques and federated learning (FL) frameworks to enhance the classification performance of medical images for different gallblader disease detection. Two different architectures were applied, which include custom Convolutional Neural Network (CNN), VGG16 models, and combining each one of these models with an FL framework.

FL is a potential strategy for training machine learning models across several sites while protecting data privacy, especially in medical image analysis. FL enables local model training, which is essential in medical fields by allowing collaboration between institutions without sharing sensitive patient data and reducing bandwidth requirements [31].

A traditional FL includes several devices working as independent clients with a single global server. Usually, a fixed number of FL and clients $\left( D \right)$ are selected randomly from a pool of edge devices that have shown their interest in participating in the learning process. For an iteration of training, each client ($d$) downloads the global learning model (${{\theta }_{t}})\text{ }\!\!~\!\!\text{ }$from the global server and start training the model with its own local data at time t. If md represents the local training dataset samples for client $d$, then $\mathop{\sum }_{d=1}^{D}=\text{ }\!\!~\!\!\text{ }md\text{ }\!\!~\!\!\text{ }=\text{ }\!\!~\!\!\text{ }M$, where $M$.  is the total size of data samples from $D$ number of clients. $f\left( \theta  \right)$ is optimized by the FL.

$f\left( \theta  \right)=\underset{d=1}{\overset{M}{\mathop \sum }}\,\frac{md}{M}{{F}_{d}}\left( \theta  \right)$                      (1)

${{F}_{d}}\left( \theta  \right)=\frac{1}{{{m}_{d}}}\underset{i\epsilon {{m}_{d}}}{\mathop \sum }\,fi\left( \theta  \right)$                            (2)

where, $fi\left( \theta  \right)$ is the loss function related with sample $i$ in the dataset of client $d$. During the local training, client $d$ updates the global learning model using an optimization algorithm like Adam and stochastic gradient descent (SGD) to minimize their loss functions. Once local model training is complete, each client sends the global server the updated global model ($\theta _{t+1}^{d}$).

($\theta _{t+1}^{d}$) = ${{\theta }_{t}}+{{\alpha }_{d}}{{\lambda }_{d}}$                        (3)

where, ${{\alpha }_{d}}$ is the learning rate and ${{\lambda }_{d}}$ is the gradient computed at client $d$ on its local data-set with ${{\theta }_{t}}$. The global server computes the improved global model $\left( {{\theta }_{t+1}} \right)$ by combining the received local models as follows.

${{\theta }_{t+1}}=\underset{d=1}{\overset{M}{\mathop \sum }}\,\theta _{t+1}^{d}$                (4)

For the next training cycle, FL clients receive the updated global model. The global model keeps going through this procedure until it converges [31].

Convolutional Neural Networks (CNNs), demonstrated excellent performance in the field of medical imaging by employing CNNs to construct an image classification model. Convolutional Neural Network (CNN) consists of 3D image input layers, convolutional layers for feature extraction, pooling layers for dimensionality reduction and compression of data, and fully connected layers to obtain distinguishable features that facilitate subsequent classification. It extracts features from images, compresses data, and learns both high-level and low-level features that are automatically extracted, eliminating the number of parameters and accelerating the system's processing [32].

Transfer learning is a process used to apply knowledge from one task to another. Feature transfer, parameter sharing transfer, and relational knowledge transfer are some of the several techniques of transfer learning. The model, like VGG16, is initialized on a pre-trained dataset and fine-tuned on the target domain. This method shortens training time, mitigates underfitting and overfitting, and enhances generalization performance. The source domain's model parameters for network initialization are taken from the ImageNet dataset. Furthermore, a new fully connected layer is built, and our dataset is used to retrain and modify every parameter in the network layers [32]. The parameter transfer approach is used in this work. This method implies pre-training a model on a different dataset (the source domain) to initialize the network and then fine-tuning the model on the GB dataset utilized for this work. This approach has the potential to reduce training time, successfully reduce underfitting and overfitting, and improve the model's generalization performance by utilizing transfer learning approaches. Accuracy (Acc), sensitivity (Sens), specificity (Spes), F1, and precision were evaluation measures. After a thorough analysis of the central server network's performance, the mean value of the best metrics. The multiclass confusion matrix evolves into an MxM matrix, where M denotes the total number of unique class labels (C0, C1, ..., CM). Their computational relationships are as follows:

$Accuracy=\frac{\mathop{\sum }_{i=1}^{N}TP\left( {{C}_{i}} \right)}{\mathop{\sum }_{i=1}^{N}\mathop{\sum }_{j=1}^{N}{{C}_{i}},j()}$              (5)

$\text{P}recision=\frac{TP\left( {{C}_{i}} \right)}{TP\left( {{C}_{i}} \right)+FP\left( {{C}_{i}} \right)}$                  (6)

$Sensitivity=\frac{TP\left( {{C}_{i}} \right)}{FN\left( {{C}_{i}} \right)+TP\left( {{C}_{i}} \right)}$                            (7)

$F1({{C}_{i)}}=\frac{2\text{*}Pre\left( {{C}_{i}} \right)\text{*}Sense\left( {{C}_{i}} \right)}{Pre\left( {{C}_{i}} \right)+Sense\left( {{C}_{i}} \right)}$                           (8)

where,

TP(C_i) is true positive for class_i, FN(C_i) is false negative for class_i.

Among them, sensitivity and specificity measures are crucial for medical picture classification tasks. However, because they contradict each other, we decided to utilize sensitivity metrics to evaluate the model's performance in our tests.

Table 3. The results of training and testing of the proposed local and global model

Dataset	Models	Acc		Loss		Pre		Recall		F1-score
		Train	Test	Train	Test	Train	Test	Train	Test	Train	Test
GB diseases Dataset	Local CNN	0.811	0.596	0.656	1.564	0.836	0.631	0.811	0.596	0.812	0.598
	Global FedAvg(CNN)	1.0	0.887	1.119	1.497	1.0	0.887	1.0	0.887	1.0	0.886
	Local VGG16	0.984	0.94	0.057	0.184	0.984	0.943	0.984	0.938	0.984	0.938
	Global Fedavg(VGG16)	1.0	0.99	0.0004	0.048	0.977	0.987	0.977	0.987	0.977	0.987

image006.png

ping_mu_jie_tu_2025-07-04_104229.png

image008.png

Figure 4. Accuracy and loss for the global FedAvg (VGG16) model and per each client

Table 4. Classification report

Class	Precision	Recall	F1 Score	Support
Cholecystitis	0.99	0.98	0.98	211
Carcinoma	1.00	0.97	0.98	211
Abdomen and retroperitoneum	0.99	0.99	0.99	181
Membranous and gangrenous cholecystitis	1.00	1.00	1.00	206
Perforation	0.99	1.00	0.99	190
Adenomyomatosis	1.00	1.00	1.00	202
Polyps and cholesterol crystals	0.95	1.00	0.98	217
Gallstones	0.99	0.97	0.98	214
Accuracy	0.99	0.99	0.99	1632
Macro avg	0.99	0.99	0.99	1632
Weighted avg	0.99	0.99	0.99	1632

image009.png

Figure 5. Multiclass confusion matrix (8 classes)

Federated Learning for the global model showed that the evaluation of the gallbladder diseases dataset revealed that most global models performed better than local models. The global FedAvg (VGG16) model achieved an accuracy of 0.997%, with a recall of 99.00%, F1-score and recall weighted average of around 98.00%, and precision weighted average of 99.00%. This model achieved a balance between avoiding false positives and identifying true positives. The global CNN model had an accuracy of 0.887%, and the global CNN model's weighted average reached 0.89%. Overall, the FedAvg aggregation method showed good performance in all global models compared to their non-federated counterparts. The Global FedAvg (VGG16) model was the most robust architecture, demonstrating perfect accuracy while maintaining privacy. Figure 4 shows the Accuracy and loss for the global FedAvg (VGG16) model and per each client after 20 rounds.

A confusion matrix is one method used to evaluate the system model's performance. When N is the number of classes, the resulting matrix is called a multiclass confusion matrix. When evaluating models that classify instances into more than two classes, this kind of confusion matrix is an extension of the confusion matrix. When dealing with problems that include more than two classes, this is quite beneficial. All of the matrix's components a class i instance's ci,j indicates how many times it was allocated to class j. The confusion matrix as a whole may have a comprehensive collection of metrics (accuracy (Acc), sensitivity (Sens), specificity (Spes), precision, and F1). The system model's performance is evaluated using a multiclass confusion matrix shown in Figure 5, with each column representing the anticipated label and each row representing the real label. Off-diagonal components indicate misclassifications, while diagonal elements indicate the number of correctly identified examples for each class, while Table 4 shows the classification report for each class.

The clinical significance of misclassifications is discussed, focusing on the impact of false positives and negatives. False positives for 'Perforation' can lead to unnecessary procedures, increased patient anxiety, and higher healthcare costs. For instance, False negatives for 'Perforation' can result in severe complications, sepsis, and even death. The model's recall for Perforation is 1.00 %, indicating no false negatives.

Figure 6 presents a comparison of model performance between the basic local learning models and global approaches of FL. Accuracy, precision, recall, and F1-score are the four assessment measures used in the comparison. Each of the models, CNN and VGG16, was tested both with and without the FL global model, allowing for a direct comparison of the effect of FL with FedAvg and pre-trained VGG16 model on system performance.

image010.png

Figure 6. Comparison of model performance between the basic transfer learning and FL approaches

In the accuracy comparison (top-left), the VGG16 model with FL achieved the highest accuracy, reaching approximately 0.99%, outperforming all other models. The CNN accuracy is 0596%, and within the global FL model, it is improved to 0.887%. The precision comparison (top-right) also highlights the superiority of the VGG16 model with FL, achieving precision close to 0.987%. The recall comparison (bottom-left) illustrates the sensitivity of the models in identifying positive instances.

The VGG16 model with FL showed the highest recall, near 0.987%. Finally, the F1-score comparison (bottom-right) shows the balance between precision and recall. The VGG16 model with FL excelled with an F1 score of about 0.987%, followed by the CNN model. Figure 6 demonstrates that FL has a significant positive impact on the performance of the CNN VGG16 model, enhancing its accuracy, precision, recall, and F1 score.