Musculoskeletal Diseases (MSDs) are the major global health concern and the second leading cause of disability worldwide. Detecting MSDs via radiography is time-consuming and difficult due to the paucity of radiologists. The traditional Convolutional Neural Networks (CNNs) have been developed to help recognize MSDs, but they are computationally expensive, requiring millions of parameters and risking data loss due to pooling processes. To overcome these issues, Aggregated-FlexiNet for Interpretable Musculoskeletal Radiograph Classification is proposed. Initially, the input images are collected from two different datasets such as MURA dataset and FracAtlas dataset. Then, Contrast-Limited Adaptive Histogram Equalization (CLAHE) technique is employed to resizing the images. Afterwards, Fractional Max-Pooling (FMP) is used to reduce the information loss with overlapping or disjoint regions. Then, Aggregated-FlexiNet is used to predict the probability for a given radiograph of a specific study as shoulder, elbow, wrist, hand, finger, humerus and forearm. The experimental results demonstrate that the Aggregated-FlexiNet achieves superior performance compared to the existing models. The proposed Aggregated-FlexiNet achieves high Cohen's Kappa value of 0.71 %, high sensitivity of 94.51 % and high specificity of 93.2 % when compared with existing models.

Musculoskeletal Disorders (MSDs) affect 1.7 billion people worldwide, causing pain and mobility issues [22]. Though convolutional neural network (CNN) models can detect MSDs, their opaque nature raise concerns regarding transparency and accountability. The class activation map (CAM) method provides insight into CNN predictions by visually highlighting regions of interest or local discrimination. This assists medical professionals in visualizing the decisions made by the CNN model. CAM utilizes activation maps of a specific convolutional layer, typically the last convolutional layer, along with the predicted class score. CAM’s local discrimination effectiveness depends on activation maps and weight computation. Therefore, CAM’s ability to local discrimination depends on the dataset and the CNN model. Therefore, it is essential to conduct an empirical study on datasets with popular CNN models used in MSD classification. This is important for ensuring unbiased decision-making in medical imaging. This study quantitatively evaluates various CAMs, including CAM, Grad-CAM, Extended-CAM, Smooth Grad-CAM and Score-CAM. Three CNN models, ResNet-50, DenseNet-169, and Inception-v3, were trained on the MURA dataset. The evaluation employs quantitative measures such as average drop, average increase, maximum coherency, and minimum complexity. Average drop coherency complexity (ADCC) is a single metric that compares qualitative measures. Results indicate that Smooth Grad-CAM and Score-CAM outperform in local discrimination regions for predictions. Smooth Grad-CAM outperforms in average drop, average increase, and ADCC for ResNet-50. However, for DenseNet-169 and Inception-v3, Score-CAM excels in ADCC.

Detection of musculoskeletal disorders is crucial for the diagnosis and treatment of related conditions, allowing timely interventions and improved patient outcomes. Despite significant advances in imaging technologies, radiologists often face challenges such as a large workload and limited resources. Although convolution neural networks (CNN) have been developed to detect musculoskeletal disorders (MSD), their performance for specific studies, such as finger and hand radiographs, often yields suboptimal Cohen kappa scores. To overcome these limitations, a CNN architecture that employs a triplet network with 128- and 256-dimensional feature embeddings is proposed to enhance classification performance for detecting finger abnormalities in MSD. Preprocessing techniques, including contrast-limited adaptive histogram equalization (CLAHE), are applied to improve the quality of input images. Advanced loss functions, such as triplet loss and focal triplet loss, are used to efficiently optimize feature embedding spaces. Subsequently, a classifier is trained using these feature embeddings to differentiate between normal and abnormal cases of MSD with high precision. This methodology offers a robust and scalable solution to support radiologists by automating abnormality detection, reducing diagnostic delays, and ultimately improving patient care and operational efficiency.

Deep learning has shown impressive results in computer vision, but it often underperforms in imbalanced datasets. This challenge is prominent in medical imaging, particularly in the classification of musculoskeletal disorders, where minority classes are underrepresented. This study presents a generative framework for the detection of musculoskeletal abnormalities with a specific focus on the problem of data imbalance in the classification of radiographic images. A disentanglement-driven approach is employed using β-variational autoencoder (β-VAE), which facilitates the generation of diverse and class-consistent samples through latent space manipulation. These synthetic samples are utilized within a triplet network for metric learning, enhancing discriminative representation by promoting greater inter-class separability and intra-class compactness in the latent space, thus effectively mitigating imbalance during classification. Experimental evaluation of the musculoskeletal radiograph (MURA) dataset, the proposed triplet network with β-VAE improves classification performance, achieving 11.3% higher accuracy and 42.3% greater Cohen’s kappa on the finger study type (DenseNet-169), and 15.6% accuracy gain with 27.1% Cohen’s kappa improvement on the forearm study type (ResNet-50), demonstrating its effectiveness for imbalance-aware diagnostic imaging.

Convolutional neural networks (CNN) are widely used to classify radiograph images. Musculoskeletal disorders (MSD) of the upper extremity (which comprises upper body parts such as the shoulder, elbow, wrist, and hand, allowing movement, strength and fine motor skills). However, their performance is often limited by class imbalance and the presence of hard samples. Although approaches like ensemble models, capsule networks and regularised CNNs in groups can address these issues, they require substantial computational resources. The adoption of loss function does not require additional computational overhead. Focal loss prioritises hard samples (samples that are not easy to classify); it simultaneously suppresses the gradients for easy samples, which affects learning. This can reduce accuracy and create an imbalance between sensitivity and specificity, which is an undesirable outcome in medical diagnostics. To overcome these limitations, adaptive binary focal loss (ABFL) is proposed here, which combines the strengths of binary cross-entropy and focal loss to achieve balanced learning between easy and hard samples. A balance parameter, (Formula presented.), is introduced to adaptively weigh the contributions of binary cross-entropy and focal loss. This approach is further extended to multi-class classification tasks through the proposed adaptive categorical focal loss (ACFL). In addition, a procedure is introduced to automatically tune the three key hyperparameters (Formula presented.), (Formula presented.) and (Formula presented.) based on the characteristics of the dataset. This eliminates the need for manual intervention. ABFL and ACFL are compared with seven existing loss functions using DenseNet-169 and Inception-v3 on musculoskeletal radiograph images (MURA), a digital database for screening mammography (DDSM) and a garbage classification dataset. Compared to focal loss, Cohen's kappa score performance improved by 33.70% in ABFL on the MURA finger dataset. Similarly, ACFL achieved improvements of 58.07% and 20.23% on the DDSM and garbage datasets, respectively, while maintaining balanced sensitivity and specificity. These results show the robustness and effectiveness of both ABFL and ACFL in handling class imbalance and hard samples in CNN-based classification.

Efficient optimization of computational resources in Mobile Edge Computing (MEC)-enabled Internet of Things (IoT) environments is critical for enhancing energy efficiency and minimizing latency. This research focuses on medical image analysis, specifically kidney tumor segmentation using CT scan images from the KiTS19 dataset. To address the challenges of high computational demands and complex feature extraction, we propose an integrated framework that combines two advanced technologies: quantum-inspired operations implemented through classical simulation along with a Residual U-Net architecture for improved computation and post-quantum cryptography for robust security. The quantum-inspired neural network design leverages principles from quantum computing to create classical algorithms that boost computational efficiency, enabling faster and more accurate segmentation of medical images within resource-constrained IoT environments. Additionally, the framework employs a quantum-inspired Double Deep Reinforcement Learning (QiDDRL) strategy to dynamically optimize resource allocation, further enhancing segmentation accuracy while reducing latency. To safeguard data transmission between IoT devices and MEC servers, the system incorporates post-quantum cryptographic techniques-specifically Kyber for encryption and Dilithium for digital signatures-providing resistance against potential quantum-computing attacks. The proposed approach achieves 98 % segmentation accuracy on the KiTS19 dataset while addressing deep learning challenges such as the vanishing gradient problem, resulting in stable model performance. Overall, this combination of quantum-inspired computing and reinforcement learning with post-quantum cryptographic protection demonstrates the potential for secure and efficient medical image analysis in IoT environments, contributing to improved diagnostics and treatment planning in resource- and security-constrained healthcare settings.

The class activation map (CAM) is useful in identifying significant image features that the convolutional neural network (CNN) model is considering while making the prediction. This is critical especially in medical diagnosis like scenarios. However, existing gradient-based methods like Grad-CAM often produce low-quality visualization results due to gradient errors despite their computational efficiency. On the other hand, non-gradient methods like Score-CAM produce quality visualization that comes with high computational costs. The proposed method SVD filters Grad-CAM (SVD-Grad-CAM), which leverages singular value decomposition (SVD) to overcome the limitations of Grad-CAM. SVD-Grad-CAM filters gradients within the gradient matrix to compute the weight of the feature map for a specific class. This filtering process is achieved by selecting the top k principal components from the SVD decomposition, which discards less important patterns and potential error data. Consequently, SVD-Grad-CAM enhances the quality of Grad-CAM by reducing the clutter of multiple region highlights. The MURA dataset, focusing on elbow study type, is utilized to assess CAM visualization quality, with a DenseNet-169 CNN model fine-tuned via transfer learning. A total of 564 validation radiographs are used in empirical comparison, showing that SVD-Grad-CAM improves average drop, average increase, maximum coherency, and Average DCC by 30%, 21.67%, 19.91% and 22.56% respectively, in comparison to Grad-CAM. Code::https://github.com/ramaiahthota02/SVD-Grad-CAM-v1.git

The generation of artificial intelligence (AI) based chest X-ray medical reports assists radiologists by saving time and improving decision-making through explainable AI techniques. Chest X-rays are extensively used to detect a variety of conditions, like pulmonary diseases and granulomatous disorders. However, existing methods for medical report generation often struggle with consistency due to their reliance on convolutional features. The proposed approach integrates gradient weighted class activation mapping (Grad-CAM) with a multilayer perceptron (MLP) and long short-term memory (LSTM) networks to produce coherent medical reports. GradCAM generates weighted localized features from activation maps of the last convolutional layer, which are ranked in descending order. These ordered features are processed by the MLP to capture feature dependencies and are subsequently analyzed by the LSTM with impressions of medical reports. The proposed model was tested on the Indiana University X-ray (IUX) dataset, which has chest X-ray images and corresponding medical reports. The impressions in medical reports are considered ground truth when training the LSTM model. EfficientNet-B0, DenseNet169 are the convolutional neural network (CNN) models that have been trained with X-rays of IUX. Integrating Grad-CAM with CNN+LSTM have significant improvement in BLEU and ROUGE scores.

In conventional architectures, passing raw convolutional neural network (CNN) feature maps directly to vision transformers (ViT) often results in suboptimal performance. This is mainly due to redundant activations and noise that can obscure relevant patterns, particularly in the case of small, localized anomalies. Additionally, directly transmitting these feature maps to the ViT limits the model’s ability to effectively capture long-range dependencies. A Fourier transform to the CNN-generated feature maps is proposed to address these limitations, converting them to the frequency domain. This transformation reduces noise and amplifies periodic patterns, thereby providing ViT with cleaner, frequency-enhanced inputs that allow for more effective self-attention across the entire image. The proposed method CNN-Fourier-ViT was evaluated on a wireless capsule endoscopy (WCE) dataset comprising 2,618 images focused on detecting bleeding versus non-bleeding conditions in gastrointestinal imaging. Experimental results demonstrated that our model outperformed traditional CNNs (e.g., ResNet-50, Inception-v3) and CNN-ViT hybrids, achieving an accuracy of 96% with a loss of 0.11. This approach illustrates the advantages of combining local feature extraction, frequency-based enhancement, and global attention for precise bleeding detection in WCE images, underscoring its potential in resource-limited clinical settings.

The Signature represents visual object shape 2D contour in 1D to recognition shape of the objectQuery. This 1D shape representation translated into Centroid Distance Histogram (CDH) Gokaramaiah et al. (Comput Graph Image Process 25:357–370, 1974 [16]) to achieve invariant transformations such as translation, scale, rotation, flip. The CDH representation performs well in content-based image retrieval system with low computational complexity and this representation method insensitive to noise of boundary. The CDH fails to represent concave shape object because the signature function maps some of the angle to more than one length from the centroid to contour. This problem solved by modifying the shape signature function which returns the average centroid length when the angle difference between two contour points approximately equals to 0.873 by traversing contour points in a clockwise direction. The starting point for clock traversing is minimum distance point from the centroid to contour. The Average Centroid Lengths (ACL) converted into histogram which makes shape representation independent of transformations. To improve recognition, more information of contour obtained by first-order and second-order difference histogram of the modified signature. This first-order and second-order difference Gokaramaiah et al. (IEEE Comput Soc, 2010 [1]) shape signature represented as hierarchical ACL. This ACL representation suitable for the Handwritten symbol recognition because small changes in the contour of shape adopted in Hierarchical ACL representation. The Handwritten symbol recognized based on k-nearest neighbor classifier (k-NNC) on sample database symbols.

Pen based inputs are natural for human beings. A hand-drawn shape (symbol) can be used for various purposes, like, a command gesture, an input for authentication purpose, etc. Shape of a symbol is invariant to scale, translation, mirror-reflection and rotation of the symbol. Moments, like Zernike moments are often used to represent a symbol. Descriptors based on Zernike moments are rotation invariant, but since they are neither translation nor scale invariant, a normalization step as pre-processing is required. Apart from this, higher order Zernike moments are error prone. The present paper, proposes to use probability distributions of some local moments of lower order, as a representation scheme. Theoretically it is shown to possess all invariance properties. Experimentally, using the k-nearest neighbor classifier (with Kullback-Leibler distance), it is shown to perform better than Zernike moments based representation scheme. © 2011 Springer-Verlag.

Content based image retrieval system (CBIR) retrieves images from a database based on the contents of the query image.Retrieval based on the shape of the 2D object present in the image is important in several applications. Shape of an objectis invariant to translation, scaling, rotation and mirror-reflection. Hence, the representation scheme which possesses all theseproperties is important. Signature histogram and k th order augmented histogram have all invariance properties [17]. But,they are applicable only to convex shapes. This representation scheme assumes that centroid to contour distance is a functionof angle (with a predefined axis). This is not true for non-convex and open shapes, since for some angles there can be more than onecentroid to contour distance. The current paper does not make this assumption, but considers distribution of centroid tocontour distances. Further, to reduce the false positive rate, distribution of local variations of the centroid contour distancesare also considered. Experimental studies are done using a standard image database and handwritten symbols database. The present technique is comparedagainst a similar recent technique. © 2011 Springer-Verlag.

Shape is an important property of visual objects which distinguishes it from others and is done with ease by living beings. Shape is invariant to scale (magnification or shrinking), rotation, translation, and flip (mirror-reflection). A good representation will lead to a better classification or retrieval system. A known representation called "normalized centroid distance signature" has all above properties except invariance to rotation and flip. By quantizing the signature into "signature histogram", invariance to rotation and flip are also achieved. But, this has a severe drawback, namely, number of false positives will increase, i.e., many entirely different shapes can have same or similar signature histograms. The paper proposes to augment the signature histogram with local information that is present at the boundary of the object. This is done in such a way without loosing any properties of the signature histogram. This leads to a hierarchical representation scheme called the "kth order augmented histogram" where k is a positive integer. A novel kth order hierarchical retrieval system is also proposed which uses the hierarchical representation scheme. The proposed scheme is compared with a recent similar scheme. © 2010 IEEE.