Journal of Machine Intelligence and Data Science (JMIDS)

, where C is the total number of channels. This ensures that the most informative channels are prioritized for subsequent refinement.

3. 2. 2. Population Diversity

To maintain robust exploration and avoid premature convergence, we define population diversity as the standard deviation of channel fitness values:

, where n is the number of channels and μ_F is the mean fitness across all channels.

A high diversity indicates that channels vary significantly in importance, signaling that exploration should continue; low diversity suggests that most channels are similar, so exploitation can be emphasized.

3. 2. 3. Fuzzy Adaptation

We employ a fuzzy inference system (FIS) to dynamically adjust the metaheuristic’s convergence parameter 𝑎, balancing exploration and exploitation:

a_k is the output from fuzzy rule 𝑘, and w_k is its weight.

Typical rules:

• IF Iteration progress is low AND diversity is high, THEN decay 𝑎 slowly → favors exploration.
• IF Iteration progress is high AND diversity is low, THEN decay 𝑎 rapidly → favors exploitation.

This adaptive control allows the algorithm to focus on different feature selection strategies depending on the current stage of training.

To ensure reproducibility and precise control, the fuzzy system is formally defined as follows: The fuzzy inference system uses two input variables: Iteration Progress (normalized epoch count ∈ [0,1]) and Population Diversity (normalized standard deviation ∈ [0,1]). Each input is mapped to two linguistic terms (“Low”, “High”) using S-shaped (for High) and Z-shaped (for Low) membership functions. The output variable, Decay Rate (a), is mapped to three terms: “Slow” (0.1–0.4), “Medium” (0.4–0.7), “Fast” (0.7–1.0). Defuzzification is performed using the centroid method. Four rules govern the system (e.g., “IF Progress=Low AND Diversity=High → Decay=Slow”). All variables are min-max normalized per epoch.

3. 3. MFO

MFO is a bio-inspired metaheuristic where moths represent candidate channels and flames represent the best channels found so far. The position of each moth is updated in a spiral pattern towards the flame:

where F_j is the flame (best channel) position, D_i=|F_j-M_i | is the distance to the flame, b=0.5 controls the spiral’s shape, and t∈[-1,1] introduces randomness. The flame counted at iteration t adapts as:

, balancing exploration and exploitation during channel selection.

3. 4. WOA

WOA is a nature-inspired metaheuristic algorithm modeled after the hunting behavior of humpback whales, particularly their bubble-net feeding strategy. In the context of feature selection for deep neural networks, WOA provides a robust method for dynamically identifying the most discriminative channels while suppressing redundant ones. The algorithm operates through three complementary mechanisms, each contributing to a balance between exploration and exploitation.

Encircling prey: This mechanism guides candidate solutions (whales) toward the best-known solution (the optimal channel) at the current iteration. It is mathematically described as:

, where represents the position of the current best channel set,is the position of a candidate channel set at iteration , and are coefficient vectors that control the step size and direction of the update, and is the distance between the whale and the best solution.

Bubble-net spiral attack: The bubble-net mechanism models the spiral movement of whales around their prey. This allows fine-grained refinement of the candidate solutions, promoting local exploitation of the feature space:

, where b is a constant controlling the spiral shape, 𝑙 is a random number in [−1,1], introducing stochasticity and represents the distance vector modified for spiral movement.

Exploration Phase: To prevent premature convergence and ensure that a diverse set of channels is considered, WOA incorporates a random search mechanism:

Here, is the position of a randomly selected channel set.

These mechanisms collectively balance exploration and exploitation, ensuring robust channel selection while preventing overfitting or redundancy.

3. 5. Feature Refinement and Aggregation

Selected channels are refined using 3×3 convolutional layers:

, where σ is the ReLU activation function.

Outputs from all relevant branches (DeepLabV3+: ASPP branches, 1×1, pooled context; U-Net: encoder layers) are concatenated:

The concatenated features are compressed using a 1×1 convolution:

This process ensures efficient integration of selected channels while maintaining spatial resolution.

3. 6. Objective Function

Training optimizes a combination of segmentation loss and channel fitness reward:

, where P is the set of selected channels, λ balances the terms, and F(C_c ) is the fitness of channel c. Maximizing F(P), ensured the selection of channels with high activations and discriminative power.

3. 7. Implementation Efficiency

Channel selection occurs only during training, minimizing computational overhead during inference. The fusion of fuzzy logic with MFO/WOA improves convergence and stabilizes selection of informative features. The framework generalizes across architectures, allowing consistent performance improvements in both DeepLabV3+ and U-Net.

4. Results and Discussion

4. 1. Datasets and Preprocessing

The first dataset comprises 135 T1-weighted MRI scans from clinically diagnosed AD patients, annotated by experienced neuroradiologists. Out of these, 100 scans were used for training and 35 for testing [22]. Patient inclusion criteria required an MMSE score ≤ 24 and hippocampal atrophy confirmed by volumetric assessment. Preprocessing included z-score normalization, intensity scaling to [0,1], and volume resampling to 1 mm³ isotropic resolution, followed by slicing into 256×224 axial slices. To simulate MRI artifacts and improve robustness, Gaussian noise (σ=0.05) was injected. The dataset was divided into training (80 subjects, 1,920 slices), validation (20 subjects, 480 slices), and test (35 subjects, 840 slices) subsets.

The second dataset is a breast ultrasound collection, widely used for classification, detection, and segmentation tasks in oncology imaging. It contains images from 600 female patients aged 25–75 years, collected in 2018. The dataset comprises 780 ultrasound images (resolution: 500×500 pixels, PNG format) along with ground truth annotations. Each image is categorized into one of three classes: normal, benign, or malignant. The availability of pixel-level masks makes this dataset well-suited for segmentation, enabling accurate tumor localization.

Both datasets highlight the challenges of medical imaging segmentation: MRI hippocampal boundaries often suffer from low contrast and tissue inhomogeneity, whereas ultrasound introduces speckle noise and variable image quality. Our framework is designed to handle both scenarios by adaptively suppressing noisy channels while retaining discriminative features.

4. 2. Implementation Details

The training was performed on two NVIDIA Tesla 4 GPUs, with the model being optimized using the Adam optimizer with an initial learning rate of 0.001 and a learning rate decay factor of 0.98 using a cosine decay scheduler. All models were trained for 100 epochs. Due to the limited memory of the GPU, the batch size was set to 16.

4. 3. Studies on Ablation

To validate the contribution of each component, experiments were conducted between the baseline DeepLabV3+, a variant without MFO (without fuzzy logic), and the proposed approach with Fuzzy-MFO and an enhanced ASPP module. The baseline model achieved a Dice score of 80.1% and an IoU of 73.11%. Adding MFO without fuzzy logic improved these metrics to a Dice score of 83.4% (improvement of 3.3%) and IoU of 75.1%. The proposed method improved further, reaching a Dice score of 85.4% (an improvement of 5.3% over the baseline) and an IoU of 77.8%. Notably, the addition of fuzzy logic enhanced the Dice score and IoU by 2% and 2.6% over the vanilla MFO by mitigating overfitting to noisy channels, respectively (p < 0.01, paired t-test) (Figure 1).

4. 4. Comparison with State-of-the-Art

To validate the effectiveness of our proposed fuzzy-optimization framework, we evaluate its performance on two medical imaging datasets: AD MRI and breast ultrasound. We compare our optimized variants, U-Net + MFO, U-Net + WOA, DeepLabV3+ + MFO, and DeepLabV3+ + WOA, not only against their respective non-optimized baselines (U-Net and DeepLabV3+) but also against state-of-the-art attention-based segmentation models, namely Attention U-Net and DA-TransUNet. Performance is assessed using Dice similarity coefficient (DSC), mean intersection-over-union (mIoU), sensitivity, and false discovery rate (FDR). Quantitative results are summarized in Table 1.

On the AD MRI dataset, our proposed method (DeepLabV3+ + Fuzzy-MFO) demonstrated consistent superiority across all evaluation metrics. It achieved the highest Dice score of 85.4%, outperforming Attention U-Net (81.22%) by 4.18% and DA-TransUNet (82.90%) by 2.5%, in addition to surpassing baseline U-Net by 6.8% and baseline DeepLabV3+ by 5.3%. For region-overlap accuracy, it reached an mIoU of 77.8%, representing a substantial gain over DA-TransUNet (75.60%) and Attention U-Net (74.31%). Sensitivity was highest at 99.7%, indicating exceptional recall of hippocampal voxels, critical for avoiding underestimation in clinical volumetric assessments. The false discovery rate decreased to 21.2%, compared to 27.0% in the DeepLabV3+ baseline and 24.5–25.8% in attention-based models, demonstrating more precise segmentation with fewer false positives. These gains confirm that our framework enhances both global structure preservation and fine boundary delineation beyond what even advanced attention mechanisms can achieve.

On the breast ultrasound dataset, characterized by speckle noise, irregular tumor boundaries, and heterogeneous textures, our U-Net + WOA variant achieved the highest Dice score of 78.80%, outperforming DA-TransUNet (78.10%) and Attention U-Net (76.87%) by 0.7% and 1.93%, respectively. It also surpassed baseline U-Net (75.75% Dice, 62.18% mIoU) by +3.05% and +4.09%, respectively. Our DeepLabV3+ + WOA} variant closely followed with 78.68% Dice} and 66.98% mIoU, surpassing its baseline (76.08% Dice, 63.48% mIoU) by +2.6% and +3.5%. Sensitivity for both top variants exceeded 98.5%, ensuring robust tumor detection, while FDR dropped to 23.0–24.0%, indicating improved localization precision over DA-TransUNet (24.8%) and Attention U-Net (26.3%). The consistent advantage over these strong attention-based baselines highlights a key strength of our approach: dynamic, fitness-driven channel selection via metaheuristic optimization provides a more adaptive and context-aware mechanism for suppressing redundancy and enhancing discriminative features than static or transformer-based attention alone.

Statistical analysis reinforced these findings. Paired t-tests (n=20 runs) indicated significant improvements (p < 0.01) for all major metrics in our optimized models compared to their respective baselines, confirming the robustness of the observed gains. Bland–Altman analysis (Figure 2: plot comparing Dice scores of baselines DeepLabV3+ vs. proposed methods across all test samples) revealed a mean bias of +5.05% with narrow limits of agreement (4.69 to 5.41), demonstrating consistent, sample-independent improvement. Collectively, these qualitative and statistical evaluations illustrate that our approach not only enhances numerical metrics but also produces clinically meaningful and visually reliable segmentations.

To evaluate the computational efficiency of our framework, we compare key metrics, including inference latency, memory consumption, and model size, in Table 2.

5. Conclusion

In this work, we presented a framework for dynamic, training-time, fitness-driven, fuzzy-adaptive channel selection using MFO and WOA. Crucially, our method enhances segmentation performance without modifying network architecture or requiring retraining, enabling transparent, plug-and-play integration with existing backbones such as U-Net and DeepLabV3+. By dynamically emphasizing discriminative channels and suppressing redundancy, guided by activation-based fitness and stabilized through fuzzy inference, the framework achieves consistent gains in accuracy and interpretability across diverse medical imaging tasks.

Extensive experiments on two clinically relevant datasets, AD MRI (hippocampal segmentation) and breast ultrasound (tumor segmentation), demonstrate the robustness and generalizability of our framework. Our best-performing variant, DeepLabV3+ + Fuzzy-MFO, achieved a Dice score of 85.4% and sensitivity of 99.7% on hippocampal segmentation, surpassing the baseline by 5.3% and 0.6%, respectively. On breast ultrasound images, DeepLabV3+ + WOA achieved 78.68% Dice, reflecting strong adaptability to noisy and heterogeneous imaging conditions. Ablation studies confirmed that the fuzzy adaptation mechanism contributes an additional 2.0% Dice improvement over vanilla MFO, highlighting its critical role in stabilizing channel selection.

Nevertheless, this study has limitations. First, our current implementation operates on 2D axial slices, which may neglect 3D anatomical context essential for volumetric consistency, particularly in longitudinal hippocampal analysis. Second, while we evaluated on two clinically relevant datasets, validation remains limited to single-center cohorts; multi-scanner and multi-institutional data are needed to assess robustness against domain shift. Third, the metaheuristic optimization introduces additional computational overhead during training, though inference remains efficient and unchanged.

Future work will address these gaps through three key directions: (1) extending the framework to full 3D volumetric segmentation to capture spatial coherence; (2) integrating it with transformer-based architectures (e.g., SwinUNETR) to combine global attention with dynamic channel selection; and (3) deploying it in federated learning settings to enable privacy-preserving model enhancement across hospitals without sharing sensitive patient data. These steps will further bridge the gap between algorithmic innovation and real-world clinical deployment.

References