ROI-guided relational YOLO–SegNet transformer for lightweight bone tumor segmentation and classification from X-ray images

Abstract

Bone tumor detection from X-ray images is challenging due to noise, low contrast, and irregular tumor boundaries that complicate precise segmentation. This study proposes a lightweight Relational YOLO–SegNet framework integrating an Optimized Savitzky–Golay Digital Filter (OSGDF), an ROI-restricted Relational Transformer Block (RTrB), and Fire Hawk Election Optimizer (FHEO)-based hyperparameter tuning. The proposed framework operates as an ROI-guided detection–segmentation pipeline, where YOLOv8 first localizes tumor regions, after which a Relational YOLO–SegNet model performs precise pixel-level segmentation. Image-level classification of normal versus tumor cases is subsequently derived from the segmented regions, making segmentation the primary objective of the framework. The unique contribution of OSGDF lies in its adaptive parameter selection using Tunicate Swarm Optimization, which improves noise suppression while preserving edge sharpness; this resulted in an SNR improvement from 21.4 dB to 29.6 dB, enhancing boundary delineation prior to segmentation. The proposed model applies relational attention only to YOLO-detected tumor regions rather than the entire image token space, reducing computational complexity while maintaining long-range contextual modeling. The framework contains 12.3 million trainable parameters, fewer than conventional encoder–decoder architectures such as UNet and Mask R-CNN. Experiments conducted on a publicly available dataset of 809 X-ray images (421 normal, 388 tumor) with expert-provided pixel-level annotations achieved 98.5% accuracy, 98.32% precision, 98.83% sensitivity (recall), 98.21% specificity, 98.57% F1-score, 97% Dice score, 97.1% Jaccard index, and an AUC of 0.981 under five-fold cross-validation (98.5 ± 0.3%). Statistical analysis confirmed that improvements over baseline models were significant (p < 0.05). The model achieved an inference time of 48 ms per image on an NVIDIA RTX 3090 GPU (24 GB VRAM), demonstrating computational efficiency suitable for resource-constrained deployment scenarios. While the results indicate strong dataset-specific performance, external multi-institutional validation is required before clinical translation. The proposed framework may serve as a potential research-support tool for automated bone tumor analysis using X-ray imaging.

Introduction

Bone tumors are abnormal growths that can affect any part of the skeletal system. Early detection is critical for improving patient outcomes and selecting the most appropriate treatment strategy¹. In clinical practice, radiologists typically rely on imaging modalities such as CT, MRI, and X-rays to evaluate suspected bone tumors. However, accurate interpretation of these images is challenging due to variations in tumor shape, intensity overlap with normal bony structures, and the often diffuse or subtle appearance of lesions. These factors may lead to delayed or incorrect diagnoses, particularly among less experienced clinicians.

In recent years, deep learning (DL)-based techniques have demonstrated significant potential in improving tumor detection accuracy owing to their ability to learn complex patterns directly from imaging data^2–4. Nevertheless, several challenges limit their effectiveness in bone tumor analysis: (i) many existing models emphasize global contextual information while insufficiently capturing fine-grained local tumor features; (ii) deep CNN architectures often involve high memory and computational requirements, restricting their suitability for real-time clinical deployment; and (iii) prior research has predominantly focused on classification tasks, with comparatively limited success in precise tumor segmentation, which is essential for surgical planning and follow-up assessment^5–7.

To address these limitations, this study proposes a lightweight DL-based segmentation framework for accurate bone tumor diagnosis^8–11. The framework adopts Relational YoLo-SegNet as the primary segmentation backbone. YoLo-SegNet is computationally efficient and lightweight while maintaining high precision for small target regions by integrating real-time object detection principles with a compact convolutional architecture^12–15. This differs from conventional backbones that rely heavily on dense encoder–decoder structures such as U-Net or DeepLab. Furthermore, a Relational Transformer Block (RTrB) is incorporated to enhance the model’s ability to capture interactions among tumor regions^16–19. By modeling long-range dependencies and inter-regional relationships, the RTrB enables the network to simultaneously exploit fine-grained local features and global contextual information, aligning with clinical requirements for precise lesion delineation.

X-ray imaging is chosen as the primary modality in this study since it is the initial diagnostic test in musculoskeletal abnormality in standard clinical practice. X-rays are cheaper than MRI and CT, available in large amounts, easier to acquire, and less radiation is emitted compared to CT; hence, they are highly applicable in mass screening and in limited resource health care facilities. X-ray is frequently the sole imaging modality in primary and rural healthcare settings with the urgent necessity to develop dependable methods of automated analysis specific to X-ray data. The interpretation of bone tumors based on X-rays is difficult by nature, as there is a lack of contrast on soft tissue, structural overlap, the inhomogeneity of the intensity, and abnormal tumor morphology. All these factors require the use of powerful preprocessing, feature extraction, and segmentation schemes. Current CNN-based systems have been effective at medical image analysis, whereas transformer-based systems have proven to be effective in modeling long-range contextual dependencies. However, full-image transformer attention makes it computationally complex, especially high-resolution medical images. Thus, the suggested solution is based on region-guided attention and optimal pre-processing to trade between the precision and the computational cost.

To further enhance performance, a Fire Hawk–Election Optimizer (FHEO) is employed. FHEO combines the local refinement capability of the Election Optimizer with the rapid global search characteristics of the Fire Hawk Optimizer (FHO)²⁰. Unlike conventional optimizers such as Adam or SGD, which may converge to local minima or require extensive manual tuning, FHEO facilitates efficient hyperparameter optimization with reduced computational overhead, thereby supporting the lightweight design of the proposed framework.

This study proposes a lightweight hybrid DL-based segmentation model to improve the detection and classification of bone tumors. The main contributions are:

• Introducing a Relational YoLo-SegNet model to accurately segment tumor regions.

• Embedding a Relational Transformer Block (RTrB) to capture fine-grained patterns and model internal relationships among tumor regions.

• Utilizing a Fire Hawk and Election Optimizer (FHEO) to tune model parameters and enhance diagnostic performance.

In this study, segmentation is considered the primary analytical task because accurate delineation of tumor boundaries is essential for clinical interpretation, surgical planning, and disease monitoring. To improve segmentation efficiency and accuracy, a preliminary detection stage using YOLOv8 is employed to localize candidate tumor regions. This ROI-guided strategy restricts transformer-based relational attention to clinically relevant areas, thereby reducing computational complexity while preserving contextual modeling. Finally, classification is performed as a secondary outcome based on the extracted segmented features to categorize the X-ray image as normal or tumor.

The remainder of this paper is organized as follows: Section "Proposed methodology" reviews related work on bone tumor detection, highlighting current challenges. Section "Results and discussion" details the proposed methodology. Section "Conclusion" presents the experimental results and comparisons with existing methods, and Sect. 5 concludes the work.

In contrast to traditional transformer-based medical segmentation frameworks that use global self-attention of full-resolution feature maps, the proposed method proposes ROI-constrained relational attention, in which transformer-based reasoning is restricted to tumor regions detected by the YOLOv8. This results in a major decrease in the amount of attention calculated whilst long-range contextual modeling is maintained in clinically pertinent spheres. Moreover, OSGDF-based edge-saving denoising and hybrid FHEO-based hyperparameter optimization are also integrated, which gives a one-stop lightweight detection-segmentation based on classification framework specifically designed to analyze bone X-rays.

Problem statement

Although the latest trends in medical image analysis based on deep learning, the current bone tumor detection and segmentation methods have various limitations, such as sensitivity to X-ray noisy image and contrast changes, underutilization of the contextual and relational features, and high computational costs, which makes them difficult to deploy effectively²¹. YOLO-based and transformer-augmented models typically do not achieve good tradeoffs between segmentation quality and low memory inference, especially when resources are limited as in clinical practice. Furthermore, standard hyperparameter optimization methods are not dynamic to the intricate search space that hybrid CNN-Transformer networks create²².

Related works

Deep learning and optimization-based techniques have been widely applied in medical image analysis for tumor detection, segmentation, and classification tasks.

Deep learning for bone tumor analysis

Rao and Madhavi²³developed BCDNet, a CNN-based model for osteosarcoma histology images, achieving 96.29% binary and 94.69% multi-class accuracy, though clinical applicability remained limited. Vaiyapuri et al²⁴. proposed HROA-DL integrating EfficientNet, BiLSTM, and a Hybrid Rice Optimization Algorithm, achieving 97.62% accuracy at the expense of high computational complexity and longer training time. Alabdulkreem et al²⁵. combined Inception v3 with Owl Search Algorithm-based hyperparameter tuning and LSTM classification, achieving 95% accuracy but reporting higher error rates. Wu et al²⁶. introduced an FCNN-CRF-based pelvic bone tumor segmentation model with a Dice score of 93.31%, although model complexity remained substantial. Xu Chen et al²⁷. utilized an improved AlexNet to classify femoral bone tumors from 500 MRI cases, achieving 98.34% accuracy with potential overfitting concerns.

Deep learning in other cancer imaging domains

Lakshmi et al²⁸. proposed a Wasserstein distribution-based constraint model (WOMT) for mammography classification to reduce false positives. Agarwal²⁹developed a CNN framework for brain tumor detection from MRI scans, where KNN classification achieved optimal performance with 99.58% across evaluation metrics. Routray et al³⁰. proposed an ensemble framework (ELSOSA-BCC) integrating EfficientNet-B0, DSAE, KELM, BiLSTM, and SOS-based hyperparameter tuning for breast cancer classification. Francis and Subramani³¹ introduced a hybrid deep learning framework for cervical cancer segmentation and classification using fractional gorilla chimp optimization-based Shepard CNN. Kollem³²proposed a four-module MRI brain tumor segmentation framework using contourlet transform and optimized SVM. Kollem et al³³. further addressed breast cancer diagnosis using recursive feature elimination and CNN feature fusion. Mounika et al³⁴. conducted a scientometric review of 348 peer-reviewed studies on MRI-based brain tumor classification, highlighting research trends and knowledge development.

Although these works demonstrate strong performance across modalities, most focus on MRI or histopathological data rather than X-ray-based bone tumor segmentation.

Optimization and image enhancement techniques

Metaheuristic optimization has been widely employed for performance enhancement. Nayak et al³⁵. applied data mining and NLP techniques for disease prediction and medicine recommendation. Palani et al³⁶. proposed an adaptive degradation-sensitive image enhancement algorithm for low-light medical images. Subramani and Veluchamy³⁷introduced a multiscale decomposition and blind denoising framework to improve structural consistency. Subramani et al³⁸. developed an optimum fuzzy intensification system using moth flame optimization for contrast enhancement. Chandrasekar et al³⁹. proposed a heart disease classification model integrating PDF-SLO-based feature selection and deep Jordan RNN. Kumar et al⁴⁰. applied EfficientNet-B3 with spatial attention for crop disease classification.

These studies illustrate the effectiveness of optimization, filtering, and enhancement strategies; however, such methods are often applied independently rather than integrated into a unified, computationally efficient bone X-ray segmentation framework.

Research gap

Despite extensive research in classification, segmentation, optimization, and enhancement across various medical imaging domains, there remains a lack of a lightweight, ROI-focused transformer-based segmentation framework specifically tailored for bone tumor analysis in X-ray images. Existing approaches either increase computational complexity or lack integrated preprocessing and optimization strategies. This study addresses this gap by proposing a computationally efficient relational attention-based segmentation architecture combined with adaptive filtering and dual-stage optimization.

Proposed methodology

In this section, a new lightweight hybrid DL based segmentation framework is discussed for segmenting bone cancer using X-ray images. The proposed system follows a three-stage pipeline consisting of (1) tumor localization using YOLOv8 detection, (2) ROI-guided tumor segmentation using the Relational YOLO–SegNet architecture, and (3) image-level classification based on the segmented tumor regions. Among these components, segmentation represents the primary task of the framework, while detection and classification serve supporting roles to improve localization and diagnostic interpretation. Initially, the raw input image is collected from publically available dataset and then pre-processed using optimized Savitzky-Golay digital filtering (OSGDF). After pre-processing, segmentation is done by utilizing Relational YoLo-SegNet based segmentation model. In order to tune the hyper-parameters for bone tumor detection with maximized accuracy, a FHEO is applied. The block diagram of the proposed work is given in Fig. 1.

Fig. 1.

Workflow of the proposed framework.

The suggested Relational YOLO-SegNet is a task specific of the available YOLO hybrid segmentation models in which relational attention is not applied to entire image token representations but only to detected tumor area, allowing capturing spatial dependencies useful to medical segmentation. In contrast to the previous methods that are based on dense feature fusion or computation-intensive global self-attention, the suggested framework uses a small relational transformer block with a smaller number of attention heads and embedding dimensions. In addition, the binding-box-first-based detection-segmentation approach, along with a small SegNet decoder, which shares encoder features, clearly indicates a focus on computational efficiency and near real-time clinical applicability, unlike the heavier transformer-enhanced segmentation networks.

Dataset

The proposed framework was evaluated using a publicly available bone tumor X-ray dataset sourced from Roboflow Universe. The dataset comprises 809 labeled X-ray images, including 421 normal and 388 tumor cases. Pixel-level segmentation masks are provided for the images and were originally annotated by domain experts, as stated by the dataset publisher. Each image contains a single ground-truth mask corresponding to tumor regions where applicable. Prior to training, all images were resized to 512 × 512 pixels and normalized to the range of 0–1 to ensure consistency and stability during model learning. The annotations provided with the dataset serve as completed reference masks; however, detailed information regarding the number of annotators and inter-annotator agreement statistics was not reported by the dataset source. Although the expert-generated annotations offer reliable supervision for training and evaluation, the absence of agreement metrics is acknowledged as a limitation of this study.

This training and evaluation of the proposed segmentation framework using the use of expert-annotated ground truth, however, offer reliable supervision. To ensure robust evaluation, the dataset was split into 70% training, 15% validation, and 15% testing sets, and five-fold cross-validation was employed to further validate the model’s generalization. Data augmentation techniques such as rotation, flipping, zooming, and contrast adjustments were applied to increase the diversity of the training data and reduce the risk of overfitting. To ensure robust performance evaluation, a five-fold cross-validation strategy was employed. The dataset was randomly divided into five equal subsets. In each iteration, three subsets were used for training, one subset for validation, and one subset for testing. This process was repeated five times so that every sample appeared in the test set exactly once. The reported performance metrics correspond to the average results across all folds, providing a reliable estimate of the model’s generalization ability.

Pre-processing

Ordinary X-ray images typically contain a variety of noises. Thereby, the standard X-ray must be smoothed and denoised using an efficient filtering technique. Recently, the Savitzky-Golay filtering algorithm (SGFA) has attracted because of its fewer parameters, rapid operating speed, as well as height and shape preservation. Typically, SGFA smoothing filter’s performance depends on a good trade-off between window size and polynomial order. It is difficult to get the optimal filtering effect with a steady window size and polynomial degree. To address this issue, an optimized Savitzky-Golay digital filtering (OSGDF) approach has been used in the proposed method of bone tumor classification to improve and smoothen the X-ray images. The tunicate swarm optimization algorithm (TSOA) and SGFA are combined in the OSGDF model^41,42. The window size and polynomial order are significantly changed by adapting the TSOA to the original SGFA⁴³. The imprecision of selecting the input filter parameters can be effectively addressed and enhanced the performance.

The proposed framework employs two metaheuristic optimization strategies with distinct functional roles: Tunicate Swarm Optimization (TSO) for adaptive Savitzky–Golay filter parameter tuning and Fire Hawk Election Optimization (FHEO) for model hyperparameter optimization. The use of two optimizers is intentional and functionally separated rather than redundant. TSO is applied exclusively during the preprocessing stage to determine optimal filter window size and polynomial order parameters that maximize signal-to-noise ratio (SNR) and edge preservation. This optimization is performed offline prior to model training and does not affect inference-time computational cost. By adaptively tuning filter parameters, the preprocessing stage improves boundary clarity and stabilizes segmentation performance across varying contrast conditions.

Classical denoising filters, such as median and Gaussian filters, are commonly used to remove noise from X-ray images but often result in blurred edges and loss of fine structural details, which are critical for detecting small or amorphous bone tumors. The Savitzky–Golay filter (SGF) addresses this limitation by smoothing noise while preserving local intensity patterns and edge sharpness, thereby improving boundary identification. However, its performance is highly sensitive to the selection of window size and polynomial degree, making manual parameter tuning challenging. To overcome this issue, the proposed Optimized Savitzky–Golay Denoising Filter (OSGDF) integrates the Tunicate Swarm Optimization Algorithm (TSOA) to adaptively determine optimal parameters. This adaptive design ensures a superior balance between noise suppression and edge preservation, enhancing tumor boundary visibility and improving the accuracy of subsequent segmentation compared to traditional fixed-parameter filtering methods.

Signal-to-noise ratio (SNR) is used as a key metric to assess how well noise reduction techniques are functioning. The following is a mathematical formulation of SNR:

1

where, characterizes the input image, suggests the denoised X-ray image of Savitzky-Golay and implies the standard deviation.

As stated earlier, polynomial order and window size are the two parameters that make the Savitzky-Golay. The following is the expression of smoothing filter:

2

3

where, characterizes the polynomial coefficients, is equivalent to , deliberates the half-width of smoothing window .

All images were denoised and normalized to the 0–1 range so that training has a uniform intensity scale. Contrast enhancement was not applied because changes to the underlying bone structure could be deliberately introduced, leading to misleading features. Flips, random rotations, and contrast jitter were applied as data augmentation methods for training to enhance the dataset diversity and the model’s ability to generalize to unseen X-ray images.

Relational YoLo-SegNet based segmentation

After pre-processing, the proposed method has used relational YoLo-SegNet model to generate a bounding box around the interested region and segment tumor region. The YOLOv8 module is used to generate the bounding box from the pre-processed X-ray images. The final segmentation images are then obtained by feeding the localized images into the Relational SegFormer network. An anchor-free model that can accurately estimate an object’s center is called YOLOv8⁴⁴. To acquire long-range dependencies, the relational YoLo-SegNet model used the BiFormer. Large model training and inference are improved by BiFormer’s use of bilevel routing, which allows for more flexible computation allocation. The enhanced YoLo-SegNet model uses a query-aware, dynamic bilevel routing attention (BiLRA). Using a global self-attention (SA) mechanism, the upper-layer router communicates with each image block and creates a global image representation. To communicate with each image block and create a local representation of image, the low-level router employs a local SA process. In BiFormer, the use of dual-layer routing attention approach improves the model’s accuracy by capturing both local and global feature information. Bone cancers are often amorphous, irregularly shaped, and scattered throughout the bone structure, making pixel-level annotation highly challenging. To address this, the proposed framework employs a two-stage pipeline. First, YOLOv8’s bounding-box-first detection strategy accurately localizes the tumor region, reducing false positives by restricting the segmentation search space to the region of interest (ROI) rather than processing the entire image. This early localization leverages YOLOv8’s strong object detection capabilities, while the subsequent SegNet decoder performs fine-grained pixel-level segmentation to refine tumor boundaries. This approach improves segmentation precision while reducing computational cost compared to direct semantic segmentation of the full image.

The proposed Relational Transformer Block (RTrB) is designed as a lightweight, task-specific adaptation of the standard multi-head self-attention mechanism. Unlike conventional transformer blocks that process full-image token embeddings, RTrB operates exclusively on ROI-extracted feature tokens corresponding to YOLO-detected tumor regions. This region-restricted design preserves relational feature modeling while significantly reducing computational overhead. Furthermore, the block is integrated within a CNN backbone to retain hierarchical feature extraction, ensuring effective compatibility with medical image segmentation tasks.

To further address the difficulty of distinguishing tumors with irregular shapes and fuzzy boundaries, a Relational Transformer Block (RTrB) is integrated into the model. By modeling long-range dependencies and contextual relationships between spatially distant regions, RTrB enhances the network’s ability to separate tumor tissue from normal bone structures and capture subtle structural cues. This leads to superior segmentation quality, particularly in cases with low contrast or minimal visual detail. Figure 2 provides the architecture of Relational YoLo-SegNet.

Fig. 2.

Architecture of Relational YoLo-SegNet.

The abbreviations used in the Relational YoLo-SegNet architecture are defined as follows:

• CBS: Convolution + Batch Normalization + SiLU activation – used as the basic feature extraction block.

• C2f.: Cross Stage Partial Network with feature fusion and channel split for efficient learning.

• BiFormer Att: BiFormer Attention mechanism that employs Bi-level Routing Attention to capture both local and global dependencies.

• K3, S2, r2: Convolutional parameters – Kernel size = 3, Stride = 2, Dilation rate = 2.

• BBox loss: Bounding Box Loss – used for object localization in detection.

• BCE loss: Binary Cross-Entropy Loss – applied for segmentation classification.

• 1 × 1 Conv: 1 × 1 Convolution – utilized for channel dimension reduction or expansion.

• Output Patch Embedding: Converts image patches into token embeddings for transformer input.

• Relational Transformer: Transformer block with relational attention for improved context learning.

• MLP: Multi-Layer Perceptron – used at the output layer for final classification.

BiLRA

A feature map is divided into non-overlapping regions such that each region encompasses . This portion is accomplished by means of reshaping as . The following is the tensor of query , value , and key , where :

4

where, the projection weights are denoted by .

Every key-value pair that are in the combination of routed areas indexed with will be attended to each query token in region . The following are the key and value tensors:

5

where, the collected key and value tensors are denoted by , and . The attention expression can be given as:

6

where, the function is parameterized with depth-wise convolution.

Relational SegFormer

Relational SegFormer is a reliable and effective segmentation approach that incorporates relational transformers with lightweight multilayer perceptron (MLP) decoder. It comprises a hierarchically structured transformer encoder. Each transformer layer’s output feature size is minimized layer by layer. In this manner, fine-grained small-scale features and coarse-grained large-scale features are extracted (i.e. capturing feature information of various scales). SegFormer⁴⁵ is a simple and lightweight decoder that can be used in place of more complicated decoder. In order to produce forceful representations, the MLP decoder combines global and local attention by combining data from several layers. As a result, it works well with bone tumor diagnosis that have significant local and global feature variances.

The dimensions of heads , , and are identical. . The following is the expression of SA:

7

The is processed using the down sampling rate in the attention calculation as follow:

8

9

where, characterizes the input’s mapping feature, represents the feature after reducing the dimension, × represents multiplication operator and represents the feature after dimension transformation.

The feed-forward network (FFN), or Mix-FFN is as follows:

10

where, indicates the feature from SA module. For every FFN, the Mix-FFN combines an MLP with a convolution. First, the channel dimension is unified in the lightweight ALL-MLP decoder part:

11

where, multilevel features are indicated by and .

Next, the features are concatenated and upsampled:

12

Then, the concatenated features are

13

Lastly, the segmentation mask can be predicted as follows,

14

where, indicates the number of categories and has resolution.

Relational transformer block: The Relation Transformer Block (RTrB), which is composed of SA and cross-attention head, is utilized to determine the dependencies between tumor regions, respectively. The query, key, and value generators , respectively, are three trainable linear embeddings in each head that are implemented via a convolution and a reshape operation. The SA head and cross-attention head pairwise query and key computations are explained as

15

16

where, the SA head is indicated by , and the cross-attention head by . The key is derived entirely from input image features in tumor regions , which is distinct from the SA head that generates the query. Instead, to incorporate adjacent structures, the cross-attention head uses the features from surrounding bone regions to generate the key.

Then the two heads’ respective attention features are calculated as follows:

17

18

Additionally, residual learning is employed to each head, yielding the following results:

19

where, a residual link of element-wise addition performs the operation and the specifies the linear embedding instigated as convolution. The structure of RTrB is given in Fig. 3.

Fig. 3.

Structure of RTrB.

Thus, the long-range dependencies are well captured by SA head, which calculates a feature’s weighted sum in each position to determine the response in that position. In bone tumor segmentation, the head is meant to further improve the edges of huge patterns and differentiate the tumor regions. The features from surrounding regions or normal bone regions are used by the cross-attention head to query global architecture.

The final output is obtained by concatenating and features from the SA head as well as the cross-attention head, respectively:

20

where, specifies the concatenation at channel dimension.

The segmented tumor regions generated by the Relational YOLO-SegNet are subsequently passed to the classification stage, where the images are labeled as either normal or tumor-affected. This ensures a clear separation between segmentation and classification tasks.

Hyper-parameter tuning using fire hawk election optimizer

The FHEO serves to optimize hyperparameters within the Relational YOLOv8-SegNet model to achieve improved segmentation performance for bone tumor classification. The FHEO is designed as a hybrid metaheuristic inspired by Fire Hawk Optimization and election-based optimization strategies, enabling a balanced trade-off between global exploration and local exploitation during the optimization process^46,47. Hyperparameter tuning plays a crucial role in deep neural network training, significantly influencing convergence, generalization, and final accuracy. Manually selecting these parameters is labor-intensive and often yields suboptimal results, especially when medical datasets are limited. Traditional optimizers such as SGD and Adam are prone to slow convergence and can easily become trapped in local minima when fine-tuning a large set of hyperparameters. Metaheuristic algorithms address this challenge by automatically exploring the parameter space to identify near-optimal configurations, thereby reducing manual trial-and-error. However, approaches like Particle Swarm Optimization (PSO) and Genetic Algorithms (GA), while effective for global exploration, often lack sufficient local search capability, resulting in suboptimal convergence. To overcome this limitation, the proposed Fused Harris Hawks and Equilibrium Optimizer (FHEO) combine the strong global exploration of FHO with the efficient local exploitation of EO. This hybrid strategy enables rapid convergence, effective avoidance of local minima, and a balanced exploration–exploitation process, making it well-suited for optimizing the numerous hyperparameters of the Relational YOLO-SegNet model, thereby enhancing both training efficiency and segmentation performance.

High-level features are extracted based on shared encoder and refined segmentation outputs, and image-level classification of normal and tumor-affected X-ray images is done. These properties are pooled through global average pooling and transferred to a low capacity fully connected classification head. In training, the classification and segmentation goals are combined with a compromise of binary cross-entropy classification and Dice segmentation, making it possible in the process to share the features effectively at a low extra computational cost.

Fire Hawk optimizer

FHO takes its inspiration from fire hawk foraging techniques because it converges quickly and has little parameter dependence while maintaining high efficiency. The core steps include:

Initialization: The candidate solutions (representing potential hyper-parameter configurations) are randomly initialized within the search space. The position of each candidate solution where, X_i is the position of the i-th candidate solution, Xmin​ and Xmax are the bounds of the search space, and rand (0,1) is a random number between 0 and 1.

21

The fitness of each candidate solution is assessed using a loss function specific to the relational YoLo-SegNet model.

Distance Calculation: The distance between each fire hawk and its nearby prey is computed to determine the influence of FHs on prey movement where Dij is the distance between the i-th fire hawk and the j-th prey, and XFH and Xpreyj are their respective positions.

22

Position Update: Fire hawks and prey positions are updated based on their interactions. Fire hawks move toward the global best solution to while prey adjust their positions to either escape, hide, or move toward safer locations where, X_best ​ is the global best solution, X_FHk​​ is another fire hawk, and X_safe​ represents a safe location within the territory.

23

The FHO is capable of extensive global exploration, which helps in preventing local minima when searching for hyperparameters, while the Election Optimizer (EO) specializes in local exploitation to fine-tune potential solutions. Both accurate refining as well as comprehensive search space coverage are facilitated through their hybridization in FHEO. This results in improved choices of learning rate, dropout, weight decay, and number of layers that immediately optimize training convergence, lower overfitting, and eventually improve the classification and segmentation accuracy of bone tumors.

Election optimizer (EO)

The EO component enhances exploitation by updating candidate positions to align with the performance observed across solutions during local refinement processes.

Opponent influence Update: Candidates adjust their strategies by observing the tactics of their opponents where X o_pponent is an opponent’s position, and α is a random number between 0 and 2 representing the degree of influence from the opponent’s strategy.

24

Mean Influence Update: Candidates refine their strategies based on voter preferences, represented by the mean value of current solutions where X_mean is the mean location of every candidate solutions in the current iteration.

25

Control Factor: A control factor is introduced to balance exploitation and exploration where t represents the current iteration, T specifies the maximum number of iterations, and β adjusts the influence of the mean position.

26

Hybrid FHEO strategy

During each iteration of the FHEO the process switches from FHO-based global search to EO-based local refinement. Both optimization methods combine to help the optimizer move past local minima yet achieve highly successful solutions. The best candidate configuration from multiple iterations will function as the final set of tuned hyper-parameters for relational YoLo-SegNet model.

FHEO is employed at the model training stage to optimize key hyperparameters, including learning rate (range: 1e⁻⁵ to 1e⁻²), batch size (8–32), dropout rate (0.1–0.5), and transformer embedding dimension (128–512). The optimizer was configured with a population size of 20 and 50 iterations, using validation Dice score as the fitness function. Convergence analysis showed stabilization within 35–40 iterations, indicating efficient search behavior.

The dual-stage optimization strategy allows independent enhancement of preprocessing robustness and model generalization without increasing inference-time latency. Since both optimizations are executed offline during training, they do not contribute to runtime computational overhead, thereby preserving the lightweight deployment characteristics of the proposed framework.

Results and discussion

The experiments were implemented in Python using PyTorch 2.2 and conducted on an NVIDIA RTX 3090 GPU with 24 GB VRAM. The model was trained for 50 epochs with a batch size of 16 and an initial learning rate of 1 × 10⁻⁴ using a cosine annealing scheduler. AdamW optimizer was employed, with binary cross-entropy loss for classification and Dice loss for segmentation.The proposed Relational YOLO–SegNet backbone contains 12.3 million trainable parameters, which is substantially fewer than conventional encoder–decoder architectures such as UNet (~ 31 M) and Mask R-CNN (> 40 M). The estimated computational complexity is 18.6 GFLOPs per inference, achieved by restricting transformer attention to YOLO-detected tumor regions instead of applying global self-attention over the full-resolution feature space. This ROI-restricted relational attention significantly reduces quadratic attention complexity while preserving long-range contextual modeling within clinically relevant regions.

The framework achieves an average inference latency of 48 ms per X-ray image (batch size = 1) and requires approximately 1.2 min per training epoch. The inference memory footprint remains below 2.1 GB, demonstrating computational efficiency suitable for near real-time research-support applications. The reduced FLOPs and memory usage compared to full-image transformer-based segmentation networks further support the lightweight characterization of the proposed architecture.

Despite the fact that the proposed framework shows good performance on the provided dataset, the rather small number of the dataset (809 X-ray images) can limit extrapolation into a variety of imaging conditions, acquisition devices and patients. The present research is aimed at confirming the efficiency of the offered design of relational attention and ROI-limited segmentation under a controlled experimental design.

Throughout each of the tests, an identical set of baseline models was kept constant to compare against. The initial group of baselines, consisting of YOLOv5, YOLOv8, Mask R-CNN, UNet, and YOLACT, was employed to compare segmentation performance. The second category consisted of BCDNet²³, HROA-DL²⁴, OSA-DL²⁵, FCNN-CRF²⁶, and the AlexNet-based model²⁷ and was used to compare overall classification and dice score performance.

Since the comparative methods reported in Table 1 were evaluated on heterogeneous hardware platforms, direct latency comparisons may introduce bias due to platform-dependent variability. To ensure a hardware-independent assessment, computational efficiency was additionally evaluated using theoretical FLOPs and parameter counts. The proposed model requires 18.6 GFLOPs per inference with 12.3 million parameters, which is lower than UNet (~ 31 M parameters, ~ 25–30 GFLOPs) and Mask R-CNN (> 40 M parameters, > 35 GFLOPs). Therefore, theoretical computational indicators such as parameter count and FLOPs are also reported to provide a more hardware-independent assessment of efficiency.

Table 1.

Inference speed comparison of proposed and existing models.

Method	Model Type	Inference Time/Image (ms)	Hardware Used
BCDNet23	Heavyweight CNN	110	CPU (Intel i7)
HROA-DL24	Hybrid DL	95	CPU + GPU (GTX 1050Ti)
AlexNet-based Model27	Lightweight CNN	70	GPU (GTX 1050Ti)
Proposed Model	Lightweight DL	48	GPU (RTX 3090)

Inference latency is inherently hardware-specific; therefore, numerical runtime comparisons across different platforms should be interpreted cautiously. The proposed framework was evaluated on an NVIDIA RTX 3090 GPU (24 GB VRAM) to provide a transparent estimate of runtime performance, while baseline methods report results under their respective experimental settings. Despite these variations, the reduced FLOPs, lower parameter count, and consistent latency demonstrate the computational efficiency and practical deployability of the proposed approach across diverse hardware environments.

The proposed lightweight model undergoes two stages: pre-processing, and lightweight segmentation. The bone X-ray images are collected from the dataset in the pre-processing stage. It might have undesirable noise. To overcome this issue, OSGDF is employed to improve the visualization of the raw X-ray images. The sample normal images collected from the employed dataset is provided in Fig. 4. From Fig. 5a, Input images are pre-processed and retrieved without any noise, as shown in Fig. 5b, and fed these pre-processed images to the next stage, which is segmentation. The tumor regions are segmented using Relational YoLo-SegNet model to improve the classification rate. Now, the improved segmented images are presented in Fig. 5c.

Fig. 4.

Sample normal images collected from the dataset.

Fig. 5.

Simulation outcomes of tumor images collected from the dataset (a) Input images (b) Pre-processing images (c) Segmented images.

Performance metrics

After segmentation, the extracted tumor regions are fed into the classification module to categorize the X-ray images as normal or tumor. The performance of this classification stage is evaluated using the following standard metrics: accuracy, precision, recall, F1-score, Dice coefficient, and Jaccard score. In the proposed method, the assessment metrics listed below are engaged to measure the superiority of proposed lightweight DL based segmentation model for bone tumor diagnosis.

Accuracy: Accuracy is calculated by dividing the total number of potential prediction outcomes by the sum of the true negatives and true positives Eq. (27).

27

where, designates the true positive, specifies the true negative, resembles the false positive and states the false negative.

Precision: The ratio of accurate positive predictions is calculated by the precision score. In the event of imbalance classes, the positive predictive value can be utilized to assess prediction success and measure quality Eq. (28).

28

Recall: Among the actual positives, recall determines correctly predicted positives. It is also recognized as sensitivity Eq. (29).

29

F1-score: The model’s performance measured by taking both recall and precision is considered as f1-score Eq. (30).

30

Dice: The dice coefficients can be stated as follows Eq. (31),

31

where, indicates the number of cases in the predicted output and characterizes the number of cases in ground truth.

Jaccard: The Jaccard coefficients is given as Eq. (32),

32

Impact of pre-processing

Table 2 shows the ablation study assessing the role of each component in the proposed framework. The absence of optimized OSGDF causes a decrease in signal quality, having a SNR of 21.4 dB, which reduces segmentation accuracy (93.2%), and Dice score (90.5%) which proves the significance of good preprocessing. Removing the relational transformer block (RTrB) degrades contextual feature modeling, causing a noticeable decline in Dice score (93.1%) despite moderate accuracy (95.6%).

Table 2.

Performance improvement through OSGDF Preprocessing.

Configuration	SNR (dB)	Accuracy (%)	Dice (%)
Without OSGDF	21.4	93.2	90.5
Without RTrB	–	95.6	93.1
Without FHEO	–	96.2	94
Full Model (OSGDF + RTrB + FHEO)	29.6	98.5	97

Without the FHEO, a hyperparameter tuning will not be optimal and, therefore, the accuracy will be less (96.2) alongside the Dice score (94.0). The OSGDF-RTrB-FHEO model offers the highest performance of 29.6 dB SNR, 98.5 accuracy and 97 Dice score with a demonstration of the complementary effect of all three.

To evaluate the robustness of the proposed model, fivefold cross-validation was performed on the dataset. The model was trained and tested five times on different data splits, and the average accuracy and Dice score were calculated. The results are summarized in Table 3.

Table 3.

Fivefold cross-validation.

Fold	Accuracy (%)	Dice Score (%)
Fold 1	98.7	97.3
Fold 2	98.1	96.8
Fold 3	98.9	97.1
Fold 4	98.3	96.6
Fold 5	98.6	97.2
Mean ± SD	98.5 ± 0.3	97.0 ± 0.3

The five-fold cross-validation results demonstrate consistently high and stable performance across all data splits. Classification accuracy ranged from 98.1% to 98.9%, with a mean of 98.5% and a low standard deviation of 0.3%. Similarly, Dice scores varied between 96.6% and 97.3%, averaging 97.0% with a standard deviation of 0.3%. The minimal variability across folds indicates stable model behavior and reduced sensitivity to data partitioning, supporting the robustness of the proposed framework.

Segmentation performance was further assessed using region-overlap metrics. The achieved Dice score of 97% and Jaccard index of 97.1% reflect strong spatial agreement between predicted masks and ground-truth annotations, demonstrating reliable tumor region delineation and consistent lesion-level segmentation performance.

Investigation with various performance indicators

This section assesses and evaluates the classification and segmentation parameters derived from the detection approaches. Figure 6 assesses the model’s recall, accuracy, f1-score, and precision score for bone tumor detection. The proposed model’s overall accuracy 98.5% surpasses the conventional techniques like UNet, Mask R-CNN, YoLOv5, YoLov8, YOLACT. But, the accuracy rate of the proposed framework for classifying the bone tumor itself cannot be considered, especially when the data is uneven and the numbers within the typical class are noticeably higher than those in other classes. Therefore, the precision and recall of each class are calculated. According to the statistics, proposed model can achieve high recall and precision values 98.83% and 98.32%, respectively. YOLACT has produced the best recall and precision results among the conventional techniques, while UNet has produced the worst. F1-score is another important measure.

Fig. 6.

Comparative assessment in terms of accuracy, precision, recall and f1-score.

The graphical representation also displays the f1-score score of the proposed framework and the existing solutions. The proposed method maintains a high f1-score 98.57%. Furthermore, from the graphical representation, it is clear that standard DL methods function differently in medical analysis when it comes to detecting bone tumors. This might be the result of a single, weakly robust model. By combining RTrB into lightweight Relational YoLo-SegNet for effective segmentation, the proposed work enhances the model’s attention and prediction accuracy by effectively capturing the intricate patterns in the image.

The segmentation performance of proposed lightweight DL method with existing segmentation techniques is exposed in Fig. 7. Comparing the proposed diagnostic system with existing approaches like UNet, Mask R-CNN, YOLOv5, YOLOv8, and YOLOACT, the proposed relational YoLo-SegNet based segmentation process has obtained good segmentation performance. When compared to alternative segmentation techniques, the relational YoLo-SegNet design performed better in terms of dice, and jaccard score. The better dice score value obtained by the proposed model for bone tumor region segmentation is 97%, whereas the existing methods such as UNet, Mask R-CNN, YOLOv5, YOLOv8, and YOLOACT have obtained 75%, 77%, 82%, 87%, and 95%. Similarly, the jaccard score achieved by the lightweight DL model is 97.1%, while the existing methods such as UNet, Mask R-CNN, YOLOv5, YOLOv8, and YOLOACT have obtained 73%, 78%, 80%, 87.5% and 96%. This resulted from the image being appropriately segmented using the incorporation of relational transformer block into YOLO-SegNet in the proposed model. The performance improvement over Mask R-CNN is primarily due to YOLOv8’s single-stage, anchor-free segmentation design, which is more robust to noise and low-contrast variations in bone X-ray images.

Fig. 7.

Comparative assessment in terms of dice and Jaccard score.

Investigation with confusion matrix

The applicability of the suggested framework to identify bone tumors was tested on 809 X-ray images, of which 421 were healthy and 388 were tumor-infested. The capability of the model to classify information is illustrated by the confusion matrix. Out of the 421 healthy samples, 416 were accurately labeled, and only five were incorrectly identified as tumors. In the same manner, from 388 tumor instances, 381 were accurately classified and 7 were misclassified as normal. The results indicate good discrimination between normal and tumor-affected bone X-rays with a total accuracy of 98.52%. Few misclassifications indicate that features are well extracted and learned, reflecting the strength and reliability of the model. Overall, the confusion matrix presents strong evidence of the model’s enhanced predictive capability and hence is appropriately suited for automated clinical bone tumor detection. The confusion matrix shown in Fig. 8 represents the combined predictions from the test sets across all five cross-validation folds, providing an aggregated evaluation of the classification performance of the proposed model.

Fig. 8.

Confusion matrix of proposed approach for bone tumor classification.

Investigation with accuracy-loss

Figure 9 displays the accuracy and loss for lightweight DL method during the training and validation stages. The attained accuracy and loss of the lightweight DL method is clearly shown in the graphical representation. During the training and validation stage, the classifiers’ accuracy and loss are evaluated by varying the size of epochs. The proposed model’s tumor detection accuracy is enhanced at the 10-epoch during the training phase. On the other hand, the proposed method’s loss decreased after epoch 15. The comparison demonstrates that the proposed method experiences a minimum loss during training and outperforms other approaches in relation to validation accuracy. The accuracy levels stabilize between the 20 and 50 epochs. Furthermore, the validation loss demonstrates that the proposed method delivers a lower loss. At epoch 15, the proposed model’s loss is lower than the other approaches, and it stabilizes between epochs 20 and 50.

Fig. 9.

Accuracy and loss analysis based on training and validation data.

Investigation with ROC

Comparative ROC analysis was used to evaluate the classification performance of the proposed framework against baseline models including UNet, YOLOv5, and YOLOv8.

As shown in Fig. 10, the proposed model achieved the highest area under the curve with a value of 0.981, followed by YOLOv8 with 0.920, UNet with 0.943, and YOLOv5 with 0.901.

Fig. 10.

Comparative ROC curves for the proposed model and baseline models.

The clear separation of the ROC curve highlights the superior ability of the proposed model to distinguish between normal and tumor cases, demonstrating its clinical effectiveness and robustness for accurate bone tumor diagnosis.

Comparison with state-of-the-art methods

To evaluate the superiority of the proposed model, its performance was compared with other state-of-the-art methods using accuracy and Dice score, as these metrics directly represent the two key aspects of our framework: classification and segmentation. Dice score reflects the spatial overlap between the predicted tumor regions and the ground truth masks, while accuracy summarizes the overall classification performance. To evaluate the effectiveness of the proposed framework, its performance was compared with several recent state-of-the-art methods: (1) BCDNet²³, a CNN-based model for bone cancer detection using histology images; (2) HROA-DL²⁴, which integrates EfficientNet and attention-based BiLSTM with hyperparameter tuning; (3) OSA-DL²⁵, which combines transfer learning with the Owl Search Algorithm; (4) FCNN-CRF²⁶, a segmentation approach that refines tumor boundaries using a conditional random field; and (5) AlexNet-based model²⁷, designed for MRI-based bone tumor classification. These methods represent diverse strategies, providing a comprehensive benchmark. As shown in Table 4, the proposed lightweight DL model outperforms these existing works, achieving an accuracy of 98.5% and a Dice score of 97%, which are higher than those of the most advanced techniques. This comparison demonstrates that the proposed framework is highly effective and suitable for accurate bone tumor detection.

Table 4.

Comparison of proposed lightweight DL method with state-of-art in terms of accuracy and dice score.

Method	Accuracy (%)	Dice score (%)
BCDNet23	96.29	92.40
HROA-DL24	97.62	94.85
OSA-DL25	98.5	91.78
FCNN-CRF26	–	93.31
AlexNet-based model27	98.34	95.24
Proposed	98.5	97

The computational runtime of the proposed Lightweight Relational YOLOv8 Transformer-based segmentation framework was evaluated during both training and inference. Experiments conducted on an NVIDIA RTX 3090 GPU (24 GB VRAM) showed a training time of 1.2 min per epoch (50 epochs total) and an average inference latency of 48 ms per image. The model contains 12.3 million trainable parameters, supporting its lightweight characterization.

Compared with existing methods, the proposed framework demonstrates superior computational efficiency while maintaining higher accuracy. BCDNet achieves 96.29% accuracy but requires 4.2 min per epoch and 50 ms per image due to its dense architecture. HROA-DL reaches 97.62% accuracy with 4 min per epoch and 48 ms inference time, though its hybrid optimization strategy increases complexity. OSA-DL (95% accuracy) requires 3.8 min per epoch and 45 ms per image. FCNN-CRF reports a Dice score of 93.31% with 3.5 min per epoch and 42 ms inference latency. An AlexNet-based model achieves 98.34% accuracy with 3.6 min per epoch and 40 ms per image; however, its conventional architecture limits computational efficiency. In contrast, the proposed model attains 98.5% accuracy and 97% Dice score with significantly reduced training time, attributed to ROI-restricted segmentation, optimized preprocessing (OSGDF), and structured hyperparameter tuning.

To validate performance superiority, paired t-tests were conducted across five-fold cross-validation results. Improvements in Dice score and classification accuracy compared to baseline models were statistically significant (p < 0.05), supporting inferential rather than purely descriptive claims. To statistically validate the performance improvement of the proposed model, paired t-tests were conducted using the Dice score and classification accuracy obtained from each cross-validation fold. The proposed model was compared with baseline methods using fold-wise performance values. A significance threshold of p < 0.05 was adopted to determine whether the performance differences were statistically significant.

Across five folds, classification accuracy ranged from 98.1% to 98.9% (mean 98.5% ± 0.3%), while Dice scores varied between 96.6% and 97.3% (mean 97.0% ± 0.3%), indicating stable and consistent model behavior. Overall performance metrics include 98.5% accuracy, 98.32% precision, 98.83% recall, 98.57% F1-score, 97% Dice score, and 97.1% Jaccard index, demonstrating strong classification reliability and lesion-level segmentation agreement. The high Dice and Jaccard values reflect substantial spatial overlap between predicted masks and ground-truth annotations, supported by the ROI-restricted relational attention mechanism.

The improved efficiency primarily results from YOLOv8-based tumor localization that constrains segmentation to clinically relevant regions, reduced attention complexity within the Relational Transformer Block, and adaptive filtering using OSGDF. FHEO-based hyperparameter optimization further improved convergence stability and minimized overfitting on the moderately sized dataset.

From a clinical perspective, the framework may serve as a research-support tool by assisting radiologists in automated tumor localization and classification, potentially reducing interpretation time and providing visual segmentation outputs to enhance interpretability. However, the experimental evaluation was conducted on a single publicly available dataset of 809 X-ray images. Although five-fold cross-validation ensured robustness, external multi-institutional or cross-dataset validation was not performed. Therefore, the reported performance should be interpreted as dataset-specific, and further validation on diverse clinical datasets is necessary to confirm generalizability in real-world applications.

Conclusion

This study proposed a lightweight ROI-guided segmentation framework for bone tumor analysis in X-ray images. The primary objective of the proposed method is accurate tumor segmentation, while detection is used to localize the region of interest and classification serves as a secondary diagnostic output derived from the segmented regions. The proposed framework combines and adapts these techniques within a computationally efficient ROI-restricted design tailored for bone X-ray analysis. Experimental evaluation on a publicly available dataset of 809 images demonstrated strong performance, achieving 98.5% classification accuracy, 97% Dice score, and 97.1% Jaccard index with statistically significant improvements over baseline models (p < 0.05). The framework contains 12.3 million trainable parameters and requires approximately 18.6 GFLOPs per inference, which is lower than conventional encoder–decoder architectures such as UNet (~ 31 M parameters, ~ 25–30 GFLOPs) and Mask R-CNN (> 40 M parameters, > 35 GFLOPs). The inference latency of 48 ms per image and memory footprint below 2.1 GB further support its computational efficiency. However, the experiments were conducted on a single dataset using five-fold cross-validation, and external multi-institutional validation was not performed. Therefore, the reported results should be interpreted as dataset-specific, and broader generalization to diverse clinical environments requires further investigation. Additionally, while the framework demonstrates near real-time inference capability, comprehensive clinical validation is necessary before deployment in routine diagnostic workflows. The proposed system has potential as a research-support tool and requires further multi-institutional validation before clinical deployment.

Author contributions

S.R. and N.C. wrote the main manuscript text and V.M.S and G.R.M prepared all figures. All authors reviewed the manuscript.

Data availability

https://universe.roboflow.com/vibhu-raj-ysy7d/bone_tumor.

Declarations

Competing interests

The authors declare no competing interests.