The clinical stakes are considerable. Reduced bone mineral density and deteriorating bone microarchitecture make weight-bearing joints fragile, and the condition poses particular problems in orthopaedic surgery. Patients undergoing total knee arthroplasty who harbor undetected low bone density face higher risks of implant instability and postoperative complications, and distal femur fractures become far harder to fix when the surrounding bone lacks structural strength. Early identification of osteoporosis from routine radiographs could therefore change how surgeons plan interventions and how physicians time preventive treatment.

Most existing AI approaches to osteoporosis screening rely on convolutional neural networks or transfer learning from pretrained image classifiers. These architectures excel at picking up local patterns, such as texture changes in trabecular bone, but they struggle to capture long-range spatial dependencies across an entire radiograph. The research team, Sudhir Kumar Sangula and Siddique Ibrahim S. P., set out to close that gap by combining several complementary mechanisms that each address a different weakness of standard convolutional models.

The pipeline begins with two blocks of depthwise separable convolutions, each followed by batch normalization, ReLU activation, and max pooling. The first block uses 64 filters and the second 128, both with 3-by-3 kernels. Separable convolutions factor a standard convolution into a depthwise step and a pointwise 1-by-1 combination, which slashes computational cost while still capturing spatial information efficiently. Input images are resized to 224 by 224 pixels and represented in three-channel RGB format even though the original X-rays are grayscale, a common convention that keeps the architecture compatible with standard deep learning frameworks.

Next comes the squeeze-and-excitation block, the first of the architecture’s attention mechanisms. Global average pooling condenses each feature channel into a single descriptor, and two dense layers with ReLU and sigmoid activations learn a set of channel weights that rescale the feature map. Channels carrying diagnostically useful information are amplified while less relevant ones are suppressed. The authors chose this lightweight channel-attention design over heavier alternatives such as CBAM or full self-attention modules because it offers a favorable trade-off between performance and computational overhead, an important consideration for medical imaging tasks with limited data.

The reshaped feature maps then flow into a transformer encoder with four attention heads and an embedding dimension of 256. By treating spatial locations as a sequence of tokens, the transformer applies multi-head self-attention to model relationships between distant regions of the image, something convolutional kernels with small receptive fields cannot do directly. Layer normalization, residual connections, and feed-forward dense layers stabilize training and deepen the contextual reasoning. The output of this stage is a set of globally informed feature embeddings that encode how bone density patterns in one part of the knee relate to structures elsewhere in the radiograph.

Two recurrent layers then refine these embeddings in sequence. A bidirectional long short-term memory network reads the feature sequence in both forward and backward directions, capturing long-range contextual dependencies, and a bidirectional gated recurrent unit follows with a more parameter-efficient gating mechanism that further polishes the representations. The authors report that stacking the two recurrent types produced consistent gains over intermediate configurations, with the BiLSTM contributing deep contextual memory and the BiGRU adding computational economy. A final softmax layer assigns probabilities to the three diagnostic classes.

Trained for 120 epochs on a publicly available Zenodo dataset of knee X-rays augmented to 5,400 images, the model reached 91.0 percent accuracy, 91.0 percent precision, 90.2 percent recall, and an F1-score of 90.4 percent on a held-out test set. Across five independent runs with different random seeds, mean precision, recall, and F1-score held steady at 0.91, 0.90, and 0.90 respectively, and the 95 percent confidence interval for accuracy spanned 89.8 to 92.2 percent. Paired t-tests confirmed statistically significant improvements over every baseline, including an ensembling method that scored 87.3 percent. Perhaps most strikingly, the entire network contains only about 1.1 million parameters and trained in under 59 seconds, compared with training times exceeding 760 seconds for VGG-16 and nearly 975 seconds for ResNet-50.

An ablation study traced the value of each component. A baseline CNN scored an F1 of 0.83; adding the SE block lifted it to 0.85, the transformer encoder to 0.87, the BiLSTM to 0.89, and the full stack with BiGRU to 0.90, with every incremental gain statistically significant. Gradient-weighted Class Activation Mapping, or Grad-CAM, provided the explainability layer, generating heatmaps that showed the network concentrating on knee joint regions, cortical bone boundaries, and surrounding bone density structures. In normal cases the activations were diffuse and mild, while osteopenic and osteoporotic images produced strong localized attention around areas of reduced bone density, suggesting the model learned clinically meaningful features rather than image artifacts.

The authors are candid about the study’s limits. Evaluation rested on a single dataset, confusion between osteopenia and early osteoporosis reflected the genuine difficulty of separating gradual bone loss on imaging alone, and the Grad-CAM interpretation remained qualitative without physician-annotated validation. They suggest that fusing complementary clinical data such as bone mineral density scores, patient age, and risk factors could sharpen discrimination between adjacent disease stages. Multi-center external validation will be essential before such a system enters clinical practice, but as a compact, fast, and transparent screening tool, the hybrid SE-Transformer architecture offers a compelling glimpse of how explainable AI could bring routine bone health assessment into the radiology workflow.

Subject of Research: Explainable hybrid deep learning for knee osteoporosis classification from X-ray images

Article Title: An explainable AI-driven hybrid SE-transformer architecture for robust knee osteoporosis classification from X-ray data

Article References: Sangula, S. K., & S. P., S. I. (2026). An explainable AI-driven hybrid SE-transformer architecture for robust knee osteoporosis classification from X-ray data. Frontiers in Signal Processing, 6, Article 1853106. https://doi.org/10.3389/frsip.2026.1853106

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1853106

Keywords: osteoporosis, knee X-ray, deep learning, transformer, squeeze-and-excitation, BiLSTM, BiGRU, Grad-CAM, explainable AI, medical imaging, computer-aided diagnosis, bone density

Researchers led by Balu P. Bhusari of the Ramrao Adik Institute of Technology, DY Patil Deemed to be University in Navi Mumbai, together with colleagues including Akshay A. Jadhav of SIES Graduate School of Technology, have developed a Dual-Path Attention-Augmented Residual Network, abbreviated DA-ResNet. The model is built on a ResNet-50 backbone, a widely used convolutional neural network architecture known for its ability to train very deep layers without suffering from vanishing gradients. What sets the new approach apart is not raw depth but a series of carefully engineered modules that address the specific physics of low Jamming-to-Noise Ratio, or JNR, conditions, where the distinguishing features of a jamming waveform are largely obscured by noise.

The problem the team set out to solve is fundamentally one of feature extraction. Radar jamming signals come in many varieties, including deceptive jamming that mimics real target returns, barrage jamming that blankets entire frequency bands, and compound jamming that combines multiple techniques. Each type leaves a characteristic fingerprint in the time-frequency domain, the two-dimensional representation of how a signal’s energy is distributed across frequency over time. When the jamming-to-noise ratio is high, these fingerprints are crisp and easy for machine learning models to read. As JNR drops toward and below zero decibels, however, the fingerprints fade into the noise floor, and models trained on cleaner data begin to fail dramatically.

To combat this, the researchers equipped their network with a Gabor-initialized convolutional front-end. Gabor filters are mathematical functions long prized in signal and image processing for their ability to capture oriented textures and periodic patterns, precisely the structures that jamming waveforms leave in spectrogram images. By initializing the first convolutional layer with Gabor filters rather than random weights, the network starts with an innate sensitivity to the directional ridges and oscillatory patterns that distinguish one jamming type from another, even when those patterns are faint. This biologically inspired initialization, which echoes the receptive fields found in the human visual cortex, gives the model a head start that random initialization cannot provide.

The second key ingredient is attention. The architecture incorporates Squeeze-and-Excitation, or SE, attention blocks, which perform channel-wise feature recalibration. In practical terms, the network learns to weigh the importance of each feature channel dynamically, amplifying the channels that carry discriminative information about the jamming type and suppressing those dominated by noise. This adaptive suppression is critical in low JNR regimes, where a substantial fraction of the raw input energy is useless or misleading. Rather than treating every feature equally, the model effectively learns which parts of the signal representation deserve scrutiny and which should be tuned out.

The third innovation is a hybrid dual-path pooling strategy at the end of the network. Most convolutional classifiers collapse their final feature maps using global average pooling, which summarizes the overall energy distribution of a signal but can wash out brief, sharp events. Global max pooling does the opposite, preserving the strongest transient peaks but ignoring the broader energy structure. The DA-ResNet combines both, running the features through parallel paths that capture global energy distribution and transient peak characteristics simultaneously. Because different jamming types differ in different ways, some through their overall spectral shape and others through sudden spikes or chirps, this dual-path design ensures that neither kind of evidence is lost before classification.

The team evaluated the model on two independent datasets containing time-frequency representations of radar jamming signals, with jamming-to-noise ratios spanning an unusually demanding range from minus 10 decibels to plus 30 decibels. The results were striking. DA-ResNet achieved an overall recognition accuracy of 97.07 percent, outperforming existing state-of-the-art methods across the tested conditions. Importantly, the model demonstrated strong cross-SNR generalization, meaning it retained high accuracy when tested on signal conditions different from those seen during training, a property that matters enormously in real deployments where the interference environment cannot be predicted in advance. The authors also note that the model achieves this performance with moderate computational overhead, an important consideration for radar systems where processing latency and hardware budgets are constrained.

Beyond raw accuracy, the researchers subjected their network to a battery of diagnostic analyses designed to confirm that it was learning meaningful, separable representations rather than exploiting artifacts. Confusion matrices showed clean separation among jamming classes with limited cross-confusion. Receiver operating characteristic curves quantified the trade-off between detection sensitivity and false alarms across classes. A t-SNE visualization, a technique that projects high-dimensional feature embeddings into two dimensions, revealed tightly clustered groups corresponding to individual jamming types, indicating that the learned features are genuinely discriminative. Perhaps most compelling for practitioners, Grad-CAM heatmaps highlighted the specific regions of the time-frequency images that the network attended to when making its decisions, offering a form of explainable artificial intelligence that builds trust in the model’s judgments and helps engineers verify that the network is focusing on physically meaningful signal structures.

The significance of this work extends beyond a single benchmark. Electronic warfare is escalating as a domain of geopolitical competition, and the ability of a radar to recognize the jamming it faces in real time determines which countermeasures it can deploy. A radar that can correctly identify a deceptive sweep jamming at minus 10 decibels JNR can switch to appropriate anti-jamming processing, such as sidelobe blanking, frequency agility, or adaptive beamforming, before the deception succeeds. Models like DA-ResNet, which remain robust deep into the noise floor, could therefore translate directly into survivability advantages for both military platforms and critical civilian infrastructure such as air traffic control and GNSS-dependent navigation, where jamming and interference are growing concerns.

The study also reflects broader trends in applied machine learning. Rather than inventing an entirely new network family, the authors combined proven components, residual learning, Gabor-based initialization, squeeze-and-excitation attention, and dual pooling, into an architecture tailored to the physics of the problem. This pattern of physics-informed deep learning, where domain knowledge shapes the inductive biases of the model, is proving especially valuable in signal processing fields where training data is limited and noise is adversarial. The researchers, who received no external funding for the work and report no competing interests, suggest that their framework’s balance of accuracy, robustness, and computational efficiency makes it a practical candidate for integration into next-generation electronic counter-countermeasure systems, where every decibel of recognition capability recovered from the noise could make the difference between a radar that sees through the storm and one that is blinded by it.

Subject of Research: Deep learning-based recognition of radar jamming signals under low jamming-to-noise ratio conditions

Article Title: Dual-Path Attention-Augmented ResNet for Robust Radar Jamming Recognition under Low JNR Conditions

Article References: Bhusari, B. P., Jadhav, A. A., Somani, S., More, S., & Patil, S. (2026). Dual-Path Attention-Augmented ResNet for Robust Radar Jamming Recognition under Low JNR Conditions. Mobile Networks and Applications. https://doi.org/10.1007/s11036-026-02550-4

Image Credits: AI Generated

DOI: 10.1007/s11036-026-02550-4

Keywords: radar jamming recognition, deep learning, ResNet-50, attention mechanism, Squeeze-and-Excitation, Gabor filters, time-frequency analysis, low JNR robustness, electronic counter-countermeasures, explainable AI, Grad-CAM, dual-path pooling