FOLC-Net: Lightweight Federated MRI Diagnosis

Updated 6 July 2026

FOLC-Net is a federated, MRFO-optimized lightweight architecture designed to improve brain tumor classification from multi-view MRI data using a ShallowFed backbone.
It integrates ConvNeXt for enhanced client adaptability and employs global model cloning to streamline decentralized training in privacy-sensitive settings.
Empirical results highlight superior accuracy—especially on challenging sagittal views—while significantly reducing parameter count and storage requirements.

Searching arXiv for papers on FOLC-Net and closely related usage. FOLC-Net is a Federated-Optimized Lightweight Architecture introduced for MRI disease diagnosis in decentralized settings, with particular emphasis on brain tumor classification across axial, coronal, and sagittal anatomical views (Khan et al., 9 Jul 2025). It was proposed to address a specific weakness reported for existing deep learning and federated learning models: strong performance on combined multi-view MRI data accompanied by substantial degradation on individual views, especially the sagittal view. In the reported formulation, FOLC-Net combines a ShallowFed backbone, Manta-ray foraging optimization (MRFO) for structural tuning, global model cloning for scalable federated training, and ConvNeXt for enhanced client adaptability, while maintaining approximately 1.217 million parameters and a storage requirement of 0.9 MB (Khan et al., 9 Jul 2025).

1. Problem setting and optimization target

The motivating problem is explicitly view-sensitive. The framework is designed for scenarios in which many hospitals or clients possess MRI data that cannot be centrally pooled because of privacy and compliance constraints, yet the diagnostic model must remain reliable both on combined multi-view data and on single-view inputs. The paper argues that multi-view training can obscure failure modes that appear when evaluation is restricted to one anatomical orientation, and it identifies the sagittal plane as the most difficult case (Khan et al., 9 Jul 2025).

The paper frames standard multi-view federated learning as

$f_{\theta} = \arg \min_{\theta} \sum_{i=1}^{n} L(f_{\theta}(D_i), Y_i),$

where each client $i$ holds decentralized data $D_i$ with labels $Y_i$ . It then emphasizes that a model optimized on aggregated views may not preserve view-specific discrimination, writing the single-view objective for the sagittal case as

$f_{\theta}^{\text{sagittal}} = \arg \min_{\theta} L(f_{\theta}(D_i^{\text{sagittal}}), Y_i).$

The stated objective is therefore to optimize performance on both the combined setting and the individual axial, coronal, and sagittal settings through

$\min_{\theta} L_{\text{view}}(f_{\theta}(D_i^{\text{view}}), Y_i).$

This formulation places FOLC-Net at the intersection of federated optimization, lightweight model design, and view-aware medical image analysis. A plausible implication is that the framework is not only a compression-oriented architecture, but also a response to a distribution-shift problem induced by anatomical orientation.

2. Architectural composition

FOLC-Net is built on a lightweight ShallowFed architecture. The described layer sequence includes input image, convolutional layers, LeakyReLU activations, MaxPooling $2\times2$ , Dropout, Batch Normalization, Flatten, a Feature Uniform / leveling block, dense layers of 128 and 64 neurons, and a final classification output (Khan et al., 9 Jul 2025). The reported hyperparameters include a first convolution with 32 filters and $3\times3$ kernels, LeakyReLU( $\alpha=0.1$ ), Dropout 0.25, BatchNorm with momentum 0.99 and epsilon 0.01, and a later dense-stage dropout of 0.5.

The paper attributes four distinct roles to the major components. ShallowFed provides a compact CNN-like backbone suitable for edge or hospital clients. ConvNeXt is incorporated to improve feature extraction and client adaptability, and is described as especially helpful in difficult views such as sagittal and coronal. MRFO performs structure tuning over architectural hyperparameters. Global model cloning replicates the same shallow global model to clients, where each client trains a local clone before synchronization (Khan et al., 9 Jul 2025).

The MRFO search space is explicitly enumerated in the paper:

Component	Possible values
Filters	8, 16, 32, 64, 128, 256, 512
Kernel sizes	3×3, 5×5, 7×7, 9×9
Activation	ReLU, LeakyReLU, Tanh, ELU
Dropout	0.1, 0.2, 0.3, 0.4, 0.5
Neurons	16, 32, 64, 128

MRFO is initialized with population = 10 candidate solutions, 8 iterative cycles, 40 epochs per cycle, and patience = 10. The paper divides MRFO into two strategy families, CYA and CHF, which it interprets as exploration and coordinated refinement, respectively. Its objective function evaluates candidate architectures by validation accuracy, retaining the best structure over iterations (Khan et al., 9 Jul 2025).

3. Federated training mechanism

The federated component uses a weighted aggregation strategy based on local data volume. The weight scaling factor is given as

$\text{Weight Scaling Factor} = \frac{|D_i|}{\sum_{i=1}^{n} |D_i|},$

after an earlier equivalent presentation through $i$ 0 and $i$ 1 (Khan et al., 9 Jul 2025). Local client optimization is written as

$i$ 2

where $i$ 3 denotes client parameters, $i$ 4 the learning rate, and $i$ 5 the local loss. Global synchronization is then performed through

$i$ 6

with $i$ 7.

The paper characterizes this as a weighted FedAvg-like synchronization. Its additional architectural claim is that cloning the Global ShallowFed model reduces computational burden, supports scalable training and early convergence, and is useful when clients have siloed view-specific data (Khan et al., 9 Jul 2025). The phrase “global model cloning” therefore refers to server-side replication of the same lightweight global model to clients, followed by local training and weighted aggregation.

The reported training hyperparameters are learning rate 0.001, epochs 30, batch size 64, optimizer Adam, and loss function categorical cross-entropy. The task is a four-class classification problem with labels glioma, meningioma, normal, and pituitary. The paper states that the standard categorical cross-entropy form is consistent with this setup (Khan et al., 9 Jul 2025).

4. Data, evaluation protocol, and measurement

The main dataset is the Kaggle Brain Tumor MRI Dataset with 7,023 MRI images, four classes (glioma, meningioma, normal, pituitary), and three anatomical views (axial, coronal, sagittal) (Khan et al., 9 Jul 2025). The evaluation is conducted on both all views and the individual view subsets. Images are resized to $i$ 8, normalized to $i$ 9, and augmented for balancing.

The reported split strategy is approximately 70% training, 10% validation, and 20% test. Beyond aggregate accuracy, the study evaluates the model separately on each anatomical orientation, which is one of the defining features of the work. The paper repeatedly contrasts this with prior studies that primarily reported performance on combined views (Khan et al., 9 Jul 2025).

The metrics are standard classification and ranking measures: $D_i$ 0

$D_i$ 1

$D_i$ 2

$D_i$ 3

The paper also uses ROC-AUC, Average Precision (AP), confusion matrices, t-SNE, chi-square test, Grad-CAM, and Grad-CAM++ (Khan et al., 9 Jul 2025).

5. Reported performance, validation, and generalization

The central empirical result is that FOLC-Net performs strongly on both combined and single-view evaluations, with the most notable gains appearing in the hardest orientation, sagittal (Khan et al., 9 Jul 2025).

Setting	Accuracy	Brief comparison
All views	98.01%	Higher than DenseNet169 at 97.40%
Axial view	99.44%	Higher than several 99.16% baselines
Coronal view	98.27%	Higher than DenseNet201 at 96.55%
Sagittal view	92.44%	Higher than NASNetLarge/MobileNetV1 at 91.86%

For the sagittal view, the paper reports class-wise precision, recall, and F1 as follows: Glioma 0.87 / 1.00 / 0.93, Meningioma 1.00 / 0.77 / 0.87, Normal 0.67 / 1.00 / 0.80, and Pituitary 0.98 / 1.00 / 0.99 (Khan et al., 9 Jul 2025). This profile is consistent with the paper’s observation that sagittal data remain the most difficult, with stronger class overlap than in axial and all-view settings.

Against the baseline ShallowFed CNN, the reported ROC-AUC and AP improvements are modest on all views and axial, but more pronounced on coronal and sagittal. For example, in the sagittal setting FOLC-Net reports AUC = 0.9842 and AP = 0.9970, compared with the baseline’s AUC = 0.9762 and AP = 0.9240. In the coronal setting, FOLC-Net reports AUC = 0.9998 and AP = 0.9970, compared with 0.9945 and 0.9795 for the baseline (Khan et al., 9 Jul 2025).

The paper also compares FOLC-Net with prior methods re-evaluated per view. On the sagittal view it reports 92.44% for FOLC-Net, versus 88.37% for Khan et al. (DL + Residual Learning), 88.95% for DL, and 90.69% for CNN-LSTM. It explicitly highlights the improvements

$D_i$ 4

$D_i$ 5

and an improvement of 1.75% over the prior best sagittal result in that table, 90.69 (Khan et al., 9 Jul 2025).

Efficiency claims are equally central. FOLC-Net is reported to use about 1.217 million parameters and 0.9 MB storage, compared with a baseline CNN at about 2.71 million parameters and 3.1 MB storage (Khan et al., 9 Jul 2025). The paper interprets this as making the model more suitable for edge or hospital-client deployment.

The reported validation extends beyond the main multiclass MRI task. On the BR35H binary MRI dataset, the baseline ShallowFed model attains 98.00% accuracy, while FOLC-Net attains 98.16%. The corresponding AUC/AP figures are 0.9880 / 0.9903 for the baseline and 0.9956 / 0.9972 for FOLC-Net (Khan et al., 9 Jul 2025). The paper further evaluates on two unseen external datasets: a histopathological dataset with Colon-ACA and Colon-N, where FOLC-Net reports 99.75% accuracy, and an MHC-CT dataset with Tumor and Notumor, where it reports 99.65% accuracy. This suggests that the learned design is not restricted to the original brain MRI benchmark.

Interpretability and statistical validation are also reported. Grad-CAM and Grad-CAM++ heatmaps are said to show more precise focus on relevant pathological regions than baseline CNNs. t-SNE is reported to show the best class separation for all views, high separability for axial, some overlap for coronal, and more overlap in sagittal, especially between meningioma and normal. The chi-square test reports 3731.3 for all views, 3800.2 for axial, 3500.4 for coronal, 3500.4 for sagittal, with DOF = 9 and $D_i$ 6 (Khan et al., 9 Jul 2025).

The term FOLC-Net is not unique across the broader arXiv ecosystem, and the medical-imaging usage introduced in 2025 should be distinguished from earlier or adjacent usages. In network coding, "Fulcrum Network Codes: A Code for Fluid Allocation of Complexity" describes Fulcrum Network Codes, also referred to in the supplied details as FOLC-Net/Fulcrum codes, a framework for low-overhead, low-complexity network coding based on an outer code over $D_i$ 7 and an inner binary code (Lucani et al., 2014). That work is unrelated to federated MRI diagnosis.

A second possible source of confusion appears in optical access networking. "Federated Learning over Next-Generation Ethernet Passive Optical Networks" introduces DWBA-FL, a Dynamic Wavelength and Bandwidth Allocation scheme for FL over 50G-EPON; the supplied details explicitly state that the paper does not introduce an acronym called FOLC-Net (Ciceri et al., 2021). The closest interpretation there is an FL-over-optical-access design, but the named contribution is DWBA-FL, not FOLC-Net.

A third nearby usage arises in connected-modes federated learning. "Federated Learning over Connected Modes" introduces Floco, while the supplied note states that the paper itself uses the name SosicFL and that the query refers to Floco / FOLC-Net as a paraphrastic label for the same core idea of learning a shared solution simplex with client-specific subregions (Grinwald et al., 2024). This is again distinct from the MRI architecture of 2025.

Within current medical-imaging usage, therefore, FOLC-Net denotes the federated, MRFO-optimized, ConvNeXt-enhanced lightweight architecture for multi-view and single-view MRI disease diagnosis introduced in 2025 (Khan et al., 9 Jul 2025). The broader naming landscape indicates that the acronym should be interpreted contextually rather than assumed to identify a single line of research.