Bird Swarm Optimized Cascaded Network Detects Tuberculosis With High Diagnostic Accuracy

A deep learning framework that integrates bio-inspired optimization with multistage image processing achieved a 99.05 percent testing accuracy in identifying pulmonary tuberculosis from chest radiographs, according to a study published in the Egyptian Journal of Radiology and Nuclear Medicine. Developing automated screening tools remains an urgent priority for global public health systems confronting the persistent toll of tuberculosis, a contagious illness caused by Mycobacterium that mainly impacts the lungs. Across 2022, approximately 10.6 million individuals received confirmed tuberculosis diagnoses worldwide, resulting in an estimated 1.6 million fatalities. Clinical presentation frequently involves non-specific systemic indicators such as fever, persistent coughing, night sweats, unexplained weight loss, chest pain, and hemoptysis, which overlap significantly with other common respiratory conditions like pneumonia. Differentiating these pathologies traditionally requires expert radiologic review of subtle infiltrates, consolidations, and pulmonary nodules, resources that remain scarce in rural and low-resource healthcare clinics. To mitigate diagnostic delays and misclassification risks, researchers developed BSCasNet, a specialized cascaded convolutional neural network whose structural hyperparameters are refined using a bird swarm optimization algorithm.

The foundational stage of the diagnostic pipeline focuses on standardizing image input and isolating clinically informative anatomical patterns while attenuating background artifacts. Standard clinical chest X-rays exhibit substantial variability in resolution, contrast, and sensor noise, complicating direct evaluation by computer vision algorithms. To establish consistency across inputs, raw radiographs with native dimensions of 512 by 512 pixels are resized to a standardized resolution of 224 by 224 pixels. The researchers subsequently apply histogram equalization to expand the dynamic range of gray-level intensities, thereby increasing the visual prominence of low-contrast pulmonary structures without compromising global tonal balance. To suppress acquisition noise without degrading structural margins, the framework employs an improved bilateral filter. In the improved bilateral filter, weights between filtering window pixels and the central pixel are assigned based on connection costs determined using the Dijkstra algorithm to identify the least expensive paths. Operating with an empirical tuning parameter calibrated to 200, the filter dynamically preserves high-gradient lung boundaries while smoothing parenchymal texture variations.

Quantitative evaluations demonstrate that this preliminary filtering mechanism substantially improves signal fidelity compared to standard medical image denoising techniques. When tested against median, mean, and Wiener filters, the improved bilateral filter achieved a mean square error of 0.9516 and a signal-to-noise ratio of 20.0701. Furthermore, the peak signal-to-noise ratio reached 48.3465, outperforming alternative filtering strategies. Following this noise reduction, the conditioned radiographs undergo automated anatomical segmentation using an adaptive fuzzy c-means clustering algorithm. The adaptive clustering formulation iteratively minimizes an objective function that balances pixel memberships against cluster prototypes through Euclidean distance calculations. By assigning fuzzy membership values, the algorithm partitions thoracic radiographs across multiple cluster granularities, separating broad regions such as the lungs and background prior to feature extraction.

Following segmentation, the computational framework extracts discriminatory second-order statistical texture attributes from isolated lung fields using gray-level co-occurrence matrix analysis. While first-order intensity histograms capture basic pixel brightness frequencies, second-order statistics quantify the spatial distribution and relative positioning of pixel pairs exhibiting specific gray levels. Texture features provide information about the structural order of an image surface and how it interacts with its surroundings, offering reliable characterisation of tuberculosis-related patterns. The algorithm evaluates the spatial relationships between reference and neighboring pixels in different orientations to extract second-order statistical texture features from normalized gray-level co-occurrence matrices. These mathematical features provide a compact, computationally efficient representation of pathological tissue irregularities without introducing the substantial information loss characteristic of linear dimensionality reduction methods such as principal component analysis, thereby supplying rich numerical inputs to the downstream neural classification network.

The classification module relies on a specialized two-phase cascaded convolutional neural network engineered to maintain high diagnostic precision while remaining significantly more lightweight than traditional deep learning benchmarks. In contrast to architectures such as AlexNet, which requires approximately 238 megabytes of parameter memory for 224 by 224 pixel inputs, the proposed cascaded architecture utilizes streamlined convolutional blocks. In Phase I, segmented patches traverse a primary backbone incorporating a convolutional layer with 16 filters of 3 by 3 dimension, followed by rectified linear unit activations to accelerate gradient convergence and mitigate local minima entrapment. The resulting feature representations undergo max-pooling subsampling via 3 by 3 kernels, followed by average pooling to smooth spatial maps and isolate representative features. Arbitrary dropout regularization is incorporated to prevent network over-reliance on individual nodes, and 1 by 1 convolutional kernels paired with a Softmax activation yield initial classification probabilities.

To overcome diagnostic ambiguities where subtle tuberculosis lesions resemble healthy pulmonary tissue, the model incorporates a second classification phase trained specifically on challenging cases. In Phase I training, the backbone model identifies ground truth tuberculosis cases as positive targets while unselected samples serve as negative references, yielding preliminary probability distributions. However, false-positive predictions and ambiguous borderline instances from this initial phase are systematically routed into Phase II as dedicated hard negative training samples. By training false-positive samples from Phase I as negative training samples in Phase II, the model improves its recognition of more challenging samples and excels at classifying unclear images. The dual classification branches are supervised using a composite cross-entropy loss function optimized via stochastic gradient descent, ensuring that the model calculates joint probability distributions across both cascade stages to accurately categorize ambiguous chest radiographs.

To maximize predictive performance without exhaustive manual trial-and-error, the researchers implemented a bird swarm algorithm to autonomously optimize the hyperparameter configuration of the cascaded convolutional network. Metaheuristic bird swarm optimization simulates the coordinated foraging and vigilance behaviors observed in avian flocks during seasonal migration, balancing broad global exploration against focused local exploitation. In the algorithmic formulation, individual candidate hyperparameter vectors function as birds within a multi-dimensional search space, with their positions updated iteratively based on personal best historical performance, global swarm discoveries, and randomized behavioral coefficients. Cognitive and social coefficients modulate foraging movements, while interference parameters dictate vigilance adjustments relative to the flock mean position. During training, validation cross-entropy loss serves as the fitness metric, with the neural network iteratively updating hyperparameter vectors until convergence is achieved at 100 iterations, yielding an optimal parameter configuration with minimal cross-entropy divergence.

Experimental validation of the BSCasNet framework was executed using a dedicated Python development environment applied to a curated benchmark dataset of 4,200 high-quality chest radiographs. The source repository comprises 3,500 healthy radiographs alongside 700 confirmed tuberculosis cases, reflecting the class imbalance characteristic of clinical epidemiological screening. To counteract class distribution skew during training, the investigators incorporated a balanced class weighting strategy. The dataset was partitioned using random sampling into 70 percent training data and 30 percent testing data, with 20 percent of the training partition reserved strictly as a validation subset for hyperparameter tuning. Consequently, 56 percent of the corpus supported direct network weight updates, 14 percent guided hyperparameter selection, and 30 percent remained completely isolated for final model evaluation. The hyperparameter optimization procedure evaluated numerous potential solutions using the bird swarm algorithm, resulting in a total training duration of 58.7 minutes.

On the isolated test subset, the trained BSCasNet system demonstrated rapid convergence, reaching a steady operating state within six training epochs and achieving a peak overall testing accuracy of 99.05 percent alongside a training accuracy of 99.10 percent. Detailed confusion matrix analysis revealed that the model correctly identified 205 positive tuberculosis cases and 1,043 normal cases, incurring only four false-negative classifications and eight false-positive errors. The diagnostic receiver operating characteristic analysis yielded an area under the curve value of 0.99, indicating exceptional discriminative separation between pathological and healthy instances. In systematic fivefold cross-validation trials, the framework demonstrated high experimental stability, recording a mean accuracy of 99.01 plus or minus 0.12 percent, sensitivity of 98.02 plus or minus 0.20 percent, specificity of 99.22 plus or minus 0.10 percent, precision of 96.24 plus or minus 0.29 percent, and an F1-score of 97.12 plus or minus 0.24 percent.

Ablation experiments progressively validated the specific contribution of each architectural component within the diagnostic workflow. A baseline convolutional neural network achieved an initial classification accuracy of 96.82 percent, which progressively improved to 97.58 percent following the integration of adaptive fuzzy c-means segmentation, and reached 98.31 percent when paired with gray-level co-occurrence texture feature extraction. Employing a cascaded dual-phase network structure elevated accuracy to 98.74 percent, while adding bird swarm hyperparameter optimization established the ultimate 99.05 percent diagnostic benchmark. Comparative benchmark evaluations demonstrated that the proposed model statistically outperformed established vision architectures, including ConvNeXt at 98.83 percent accuracy and Swin Transformer at 98.67 percent accuracy. Independent evaluations on external datasets confirmed generalizability, correctly identifying 323 normal and 331 tuberculosis cases on the Shenzhen cohort and 78 normal and 57 tuberculosis cases on the Montgomery repository. Despite these robust numerical findings, the authors emphasized several operational limitations, noting that the study evaluated binary classification rather than differentiating multi-class respiratory conditions like pneumonia, involved a relatively modest sample size, lacked prospective clinical trials, and required computational overhead that may challenge real-time edge deployment in austere clinical settings.

Subject of Research: Automated tuberculosis diagnosis using bird swarm algorithm optimization and cascaded convolutional neural networks

Article Title: Optimised cascaded CNN for tuberculosis diagnosis: integrating bird swarm algorithm with advanced image segmentation

Article References: F, J. S. K., F, J. S. K., V, M., & R, R. (2026). Optimised cascaded CNN for tuberculosis diagnosis: integrating bird swarm algorithm with advanced image segmentation. Egyptian Journal of Radiology and Nuclear Medicine, 57(1), Article 193. https://doi.org/10.1186/s43055-026-01854-5

Image Credits: AI Generated

DOI: 10.1186/s43055-026-01854-5

Keywords: tuberculosis diagnosis, chest radiography, deep learning, convolutional neural networks, bird swarm algorithm, image segmentation, adaptive fuzzy c-means, bilateral filter, grey-level co-occurrence matrix, feature extraction, metaheuristic optimization, cascaded neural network, medical image processing, computer-aided detection, Shenzhen dataset, Montgomery dataset, diagnostic imaging