International Journal of Electrical and Computer Engineering (IJECE) Vol.
No.
October 2025, pp.
ISSN: 2088-8708.
DOI: 10.
11591/ijece.
Bone-Net: a parallel deep convolutional neural network-based bone fracture recognition Md.
Hasan Imam Bijoy1,2.
Nusrat Islam Kohinoor3.
Syeda Zarin Tasnim3.
Md.
Saidur Rahman Kohinoor3,4 Health Informatics Research Lab (HIRL).
Daffodil International University.
Dhaka.
Bangladesh Department of Computer Science and Engineering.
Daffodil International University.
Dhaka.
Bangladesh Department of Computer Science and Engineering.
Leading University.
Sylhet.
Bangladesh Department of Information and Computer Science.
King Fahd University of Petroleum and Minerals.
Dhahran.
Saudi Arabia
Article Info
ABSTRACT
Article history:
Many people suffer from bone fractures, which can result from minor accidents, forceful blows, or even diseases like osteoporosis or bone cancer.
In the medical realm, accurately identifying bone fractures from X-ray images is paramount for effective diagnosis and treatment.
To address this, a comparative study is conducted utilizing three distinct models: a traditional convolutional neural network (CNN).
MobileNet-V2, and a newly developed parallel deep convolutional neural network (PDCNN).
The primary aim is to evaluate and contrast these models in terms of precision, sensitivity, and specificity for diagnosing bone fractures.
X-ray images of fractured and non-fractured bones are sourced from Kaggle and subjected to various image processing techniques to rectify anomalies.
Techniques such as cropping, resizing, contrast enhancement, filtering, and augmentation are applied, culminating in canny edge detection.
These processed images are then used to train and test models.
The results showcased the superior performance of the newly developed PDCNN model, achieving an impressive accuracy of 92.
89%, surpassing both the traditional CNN and pretrained MobileNet-V2 models.
A series of ablation studies are conducted to fine-tune the hyperparameters of the PDCNN model, further validating its Throughout the investigation.
PDCNN consistently outperformed MobileNet-V2 and traditional CNN, underscoring its potential as an advanced tool for streamlining bone fracture identification.
Received Mar 23, 2024 Revised May 25, 2025 Accepted Jun 30, 2025 Keywords:
Bone fracture Canny edge detection Convolutional neural network MobileNet-V2 Parallel deep convolutional neural network This is an open access article under the CC BY-SA license.
Corresponding Author:
Md.
Hasan Imam Bijoy Department of Computer Science and Engineering.
Daffodil International University Daffodil Smart City.
Dhaka-1216.
Bangladesh Email: hasan15-11743@diu.
INTRODUCTION
A mineral made of calcium is called bone .
Human bone provides a mechanical structure for the human body .
It maintains calcium homeostasis and substantial storage of phosphate, magnesium, potassium, and bicarbonate.
Human bone also helps in muscle activities .
One of the most interesting natural composite materials is human cortical bone, which acts as a structural foundation that can withstand injury and allows the human body to self-repair.
The mechanical characteristics of bone enable it to withstand fracture under diverse physiological loading conditions .
A few of these conditions that lead to bone diseases are vitamin-D insufficiency, osteoporosis.
PagetAos disease of the bones, and many more.
The amount and quality of bone tissue in the human skeleton start to decline after reaching peak bone mass in the Journal homepage: http://ijece.
Int J Elec & Comp Eng
ISSN: 2088-8708
third or fourth decade of life .
Among the most common injuries in todayAos world are bone fractures.
About 8.
9 million people experience this disorder each year .
, and the consequences of leaving a fracture untreated may end in death or permanent damage by several types of bone fracture as shown in Figure 1.
The doctors bear a great deal of responsibility for this, as they assess lots of X-ray images daily basis.
The majority of the technology used for the initial diagnosis is X-rays, a modality that has been around for more than a century and is still widely used.
Medical professionals find it difficult to assess X-ray images for many reasons, like the fact that X-rays may obscure certain unique characteristics of the bone or that it takes extensive training to accurately identify the various kinds of fractures.
A precise classification of the fracture among standard types is crucial for both the future outlook and the effectiveness of the treatment .
Figure 1.
Several types of fracture in bone Many advances in technology are making it easier to diagnose bone fractures in today's world.
Automated image processing models, such as artificial intelligence (AI), deep learning (DL), and machine learning (ML), are quick and reliable for identification, localization, and classification .
Determining the precise location of a fracture and its degree of impact might be beneficial.
The discipline of computer-aided diagnosis is an emerging field of research where computer technologies are used to offer prompt and precise It uses X-ray images and preprocesses them as needed to detect fractures in the bones .
They are getting better with each passing day as a result of learning from labeled data .
Several works have been discussed, including the following.
Luis and Ruano .
suggested a computer-aided system for bone fracture detection.
Although X-ray pictures are typically used to diagnose bone fractures, and suggested an approach that uses computers to help detect bone fractures.
Achieved 89% classification accuracy by utilizing a variety of techniques for fracture recognition, bone line detection, and speckle reduction.
It is limited to stress fractures, and this computer-aided approach is unable to classify complex fractures accurately.
In 2015.
Anu et al.
presented work using preprocessing techniques to reduce noise from the photos and transform the RGB X-ray images to grayscale using a median filter.
They then detected the edge of the image by using the Sobel edge detector.
They used the gray level co-occurrence Matrix (GLCM) to extract the feature.
Finally, the data were divided into fractured and non-fractured categories using a variety of classifier types, including decision trees (DT), neural networks (NN), and metaclassifiers.
There are 40 photos in the collection, 20 of which have fractures and 20 of which do not.
They achieved the best accuracy in the meta-classification algorithm, at 85 percent.
Tripathi et al.
stated that the location of small or hairline femur fractures is the main topic of this paper.
To determine whether or not there is a fracture, they employed SVM.
30 X-ray images compose the dataset.
The logarithmic operator is employed to enhance photos, and median and average filtering are used to eliminate noise from images.
Morphological operations and the sobel edge detection approach are employed for feature extraction.
They 7 percent accuracy in classifying the data using support vector machines (SVM) into fractured and non-fractured categories.
Sinthura et al.
suggested a convolutional neural network (CNN) to detect bone fractures.
They took an x-ray image as input, then preprocessed it using a median filter, and the preprocessed image was applied to the discrete wavelet transforms (DWT) stage, which is used to find the edge in each channel.
Then the output is compared with the database with neural networks, which gives the Vasilakakis et al.
suggested fuzzy phrases (FP) for detection.
This work aims to use the wavelet fuzzy phrases (WFP) approach to identify bone fractures using x-ray pictures.
The accuracy of the classifying approach was 84%.
Yadav and Sandeep .
built a deep CNN model to actively classify fractured or healthy After augmentation, the dataset had a size of 4000, where 2000 is a healthy bone and 2000 is cancerous, they used five-fold cross-validation and have the best accuracy of all, scoring 92.
44 percent.
Bone-Net: a parallel deep convolutional neural network-based bone A (Md.
Hasan Imam Bijo.
A ISSN: 2088-8708 Rao et al.
proposed a method to detect bone fractures.
Their dataset had 300 X-ray pictures, and they achieved 90% accuracy by using a back propagation neural network (BPNN) and SIFT feature extraction.
Karimunnisa et al.
suggested a model using 900 X-ray pictures, of which 400 were normal and 500 were fractured.
Initially, input X-ray pictures are transformed into grayscale images.
Their BPNN provides an improved classification rate of 91 percent.
Pathare et al.
employ many processing approaches, including segmentation, edge detection, and feature extraction.
Bekkanti et al.
suggested the Harris corner detection technique, which is a traditional computer vision technique related to feature detection and traditional image processing to detect bone fractures.
There were two hundred non-fractured and three hundred fractured X-ray images in their collection.
The input x-ray images are preprocessed using m3 filtering.
segmented using canny edge detection, features extracted using Harris corner detection, and lastly categorized as fractured or not.
They contrasted their efforts using SURF with BPNN and MLP-based BPNN.
While SURF using BPNN yields an accuracy of 85 percent.
MLP-based BPNN yields an accuracy of 85 percent, and their suggested approach yields a 94 percent accuracy.
In 2022.
Samothai et al.
presented advanced CNN YOLO models in this work, where the widely used YOLO-X and YOLO-R models were implemented.
With 76 percent confidentiality.
YOLO-X can locate fractures with great accuracy.
However.
YOLO-R exhibits inaccurate fracture location detection.
Kosrat and Hawezi .
analyze various ML classifiers and assist surgeons in accurately diagnosing bone fractures.
They employed a filtering technique to remove noise and preprocessing to convert the RGB images to grayscale.
In this study.
SVM is found to be the best.
Certain issues persist in the current state of bone fracture detection, necessitating improvements in accuracy to enhance detection efficacy, as outlined in Table 1 and the aforementioned study.
The augmentation of the dataset is imperative, requiring the inclusion of high-quality photos.
This study utilizes a dataset comprising a substantial number of bone fracture images, implementing canny edge detection during preprocessing to optimize results.
Despite the growing accessibility of machine-aided bone fracture detection, accuracy remains a concern, often leading to false positives.
The identification of fractures in the absence of actual occurrences is a prevalent problem, attributed to the limited volume of data and inaccuracies in the labeling of training data.
Addressing these challenges, this research focuses on refining training and testing data by introducing a newly customized parallel deep learning model with parameter customization.
Notably, publicly available datasets frequently lack flawless preprocessing and accurate labeling, necessitating the incorporation of additional labeled data to augment inputs for the model.
To ensure an adequate supply of high-quality data, edge detection algorithms are employed during data preprocessing.
The contributions of this study can be summarized as follows:
Correcting anomalies in the dataset through several image processing technique involves employing a range of image processing techniques, such as cropping, resizing, contrast enhancement, filtering, augmentation, and culminating in canny edge detection.
The study introduces a novel and tailors parallel deep convolutional neural network (PDCNN) model, designed to surpass the existing literature by enhancing accuracy in bone fracture detection.
A comparative evaluation is conducted among three CNN-based modelsAiCNN.
MobileNet-V2, and the newly proposed PDCNN.
Four cases of ablation studies is carried out to validate the proposed PDCNN The objective is to identify the most effective model for bone fracture recognition through a thorough analysis of their performance.
Table 1.
Summarized literature review of bone fracture recognition using machine learning models Study .
Data
Preprocessing Technique Canny edge detection Model
SNAKE
Sobel edge detector
Sobel edge detector
discrete wavelet transforms (DWT) Meta
Classifier SVM
CNN
M3 filter Canny Edge Detection and Conservative Filtering .
Canny edge detection NM: Not mentioned Accuracy Limitations/Remarks Only detects stress fractures and is unable to classify complex fractures accurately.
There are not many images in the dataset.
The dataset was not good.
The dataset was not good.
The dataset was not given.
WFP
SIFT
BPNN
BPNN
HTM
SVM
This was the early stage of fuzzy phrases.
development is required before they can be reliably detected.
HTBFD.
ANN, and other algorithms are useful tools for improving fracture placement.
The dataset contains many types of bone shapes, which may become confusing.
An X-ray of the skull, pelvis, and spinal column was not able to be obtained.
It only compared the models.
Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4692-4704
Int J Elec & Comp Eng
ISSN: 2088-8708
RESEARCH METHOD
This study centers on the detection of bone fractures in X-ray images through the utilization of a CNN-based model.
The development of the proposed system involves sequential steps, including data collection, image pre-processing, and fracture recognition using CNN models.
The schematic representation of the working process is outlined in Figure 2.
Figure 2.
The workflow diagram of proposed study Dataset description This study makes use of a public dataset of bone fracture and non-fractured x-ray images.
To keep things simple, the publicly available dataset of binary-class bone fracture and non-fracture x-ray images was obtained using the Kaggle platform.
for the sake of this paper, this data is referred to as the Aubone fracture datasetAy .
It is employed in the creation of an image classifier that identifies bone fractures in given x-ray There are 1899 photos identified in the testing class .
racture and non-fractur.
and 8,884 images in the training class .
racture and non-fractur.
The images were not the same size, and some of them included anomalous data in addition to being unclear.
The sample image of fractured bone and fresh bone of x-ray image is given in Figure 3.
Figure 3.
represents the fractured bone, while Figure 3.
depicts the nonfractured bone.
Figure 3.
The sample image of .
fractured bone, and .
non-fractured bone Data preprocessing In the realm of bone fracture identification, a crucial step in the research applications of computer vision and image processing involves image preprocessing.
This process serves to enhance the quality of images, reduce noise levels, rectify distortion, and prepare the images for subsequent analysis .
The following are key strategies .
ee outcome of Figure 4.
) employed in this study, along with Cropping In bone fracture identification, cropping is a prevalent image processing technique utilized to extract the region of interest (ROI) within an image .
This facilitates the isolation of relevant anatomical structures for more accurate analysis, disregarding unnecessary elements.
The output of cropping from the raw picture of Figure 4.
, can be seen in Figure 4.
Bone-Net: a parallel deep convolutional neural network-based bone A (Md.
Hasan Imam Bijo.
A ISSN: 2088-8708 Resizing Resizing plays a vital role in standardizing the dimensions of images within a dataset.
This ensures uniformity, facilitating efficient processing and analysis .
It also aids in accommodating variations in image resolutions commonly encountered in medical imaging.
Resizing is employed to standardize the dimensions of images, ensuring uniformity across the dataset with 227y227 pixels and the output shown in Figure 4.
Enhancing contrast Enhancing contrast is crucial to accentuate subtle details within the images .
This allows for better differentiation between healthy and fractured bone structures, enhancing the overall interpretability of the images.
After enhancing the image quality.
Figure 4.
displays the image with improved visual clarity.
Filtering Filtering is employed to emphasize relevant structural details while suppressing noise.
selectively enhancing certain image characteristics, filtering contributes to the improvement of image clarity and the extraction of essential information .
After filtering and enhancing the essential features, the image is displayed in Figure 4.
Canny edge detection Canny edge detection (CED) is an advanced image processing technique used in computer vision to identify edges and boundaries within an image .
It was developed to address the challenges of edge detection by providing accurate and reliable results while minimizing false positives.
In the context of bone fracture recognition using computer vision and image processing, canny edge detection plays a crucial role in highlighting prominent contours and edges within medical images.
This technique enhances the visualization of structural details in bone images, making it particularly valuable for pinpointing fractures and irregularities in the bone structure which is displayed in Figure 4.
Augmentation Augmentation is crucial for enhancing the robustness of bone fracture identification models.
exposing the model to diverse orientations, scales, and perspectives, it becomes more adept at accurately identifying fractures under a range of conditions, contributing to improved generalization and performance .
In this study, some primary image augmentation method is used such as rotation, flipping, zooming, and shearing and the outcomes of the filtered images and CED are displayed in Figure 4.
and Figure 4.
After undergoing extensive image processing, the raw image is subjected to contrast enhancement, as illustrated in Figure 4.
The dataset comprises more than 10, 783 images categorized into fractured and non-fractured classes, exhibiting a near balance between the two.
Following an 85:15 split for training and testing, the dataset consists of 8884 images in the training set and 1899 images in the test set.
Figure 4.
Several image processing applied into .
raw image and getting .
cropping image, .
resize image, .
enhance contrast, .
filtering, .
augmentation on filtered image, and .
canny edge detection and .
augmented CED image to increase the size of the dataset Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4692-4704
Int J Elec & Comp Eng
ISSN: 2088-8708
Model implementation Following the preprocessing phase, this study endeavors to construct a bone fracture recognition system by employing CNN based models.
In pursuit of high accuracy, we explore a variety of models including the traditional CNN model, a CNN-based transfer learning model.
MobileNet-V2, and our novel PDCNN model.
CNN.
MobileNet, and the newly proposed PDCNN exhibit remarkable efficacy in identifying bone fractures from X-ray images.
The succinct description of each model is given below:
Convolutional neural network Convolutional neural networks .
, or CNNs, are widely utilized in deep learning network models and computer vision algorithms.
Because it can identify patterns in images, this kind of artificial neural network is utilized for image recognition and processing.
Convolutional, pooling, and fully connected layers are some of the layers that make it up.
The convolutional layer, which makes up the majority of CNN, is where characteristics like forms, edges, and patterns are extracted from the input image by applying filters.
One or more fully connected layers are then applied to the output of the pooling layers to classify or predict the image.
MobileNet-V2 MobileNet-V2 is a lightweight convolutional neural network designed for mobile and embedded It improves efficiency with depth wise separable convolutions and inverted residuals with linear bottlenecks .
It also incorporates expansion and squeeze-excitation modules for better feature learning.
With its streamlined architecture.
MobileNet-V2 achieves high accuracy while minimizing computational resources, making it ideal for mobile applications requiring fast and efficient image processing.
Parallel deep convolutional neural network (PDCNN) A parallel architecture with two convolutional neural networks (PDCNN.
is presented to handle the recognition and classification of bone fractures in X-ray pictures.
The suggested architecture entails the following sequence of events: X-ray images with bone fractures are fed into the PDCNNs .
input layer.
To reduce the computational complexity, these images are preprocessed.
The input images have been normalized to a 277y227-pixel resolution for training, allowing for differences in pixel widths and heights.
The subsequent procedure that helps to simplify complexity is to convert the input images to canny edge The architecture of the PDCNN is then used to classify input X-ray pictures by combining, output, local, and global path.
The SoftMax function is used in the output pathway to carry out the classification of bone fractures.
Figure 5 shows the PDCNNAos structure.
Figure 5.
The diagram of proposed PDCNN model that contains four stages which are local path, global path, merging them, and output stages in order to identify the bone fracture from the inputed x-ray image The proposed PDCNN model for bone fracture identification integrates various layers to facilitate accurate recognition of fractures in X-ray images.
Initially, 2D convolutional layers are employed to extract pertinent features from the input images.
Subsequently, rectified linear unit (ReLU) layers introduce nonlinearity to the network, enhancing its ability to capture complex patterns.
A 2D cross-channel normalization layer normalizes activations across channels, contributing to improved model performance.
Following this, 2D max pooling layers reduce the dimensionality of feature maps while retaining essential information .
Concatenation merges features learned from different pathways within the network, enriching the Bone-Net: a parallel deep convolutional neural network-based bone A (Md.
Hasan Imam Bijo.
A ISSN: 2088-8708 representation of fracture-related features.
Batch Normalization is applied to accelerate and stabilize the training process, leading to faster convergence and improved generalization.
Fully connected layers process high-level features extracted by convolutional layers, enabling the network to generate predictions for fracture identification.
Dropout regularization is utilized to mitigate overfitting by randomly dropping units during training.
Finally.
Softmax activation facilitates robust classification decisions by allowing the network to predict classes based on features extracted through various parallel paths.
Table 2 presents the specifications of each layer in the proposed PDCNN model, including the layer type, properties, activation function, learnable property, and number of learnable parameters.
Table 2.
The specifications of proposed PDCNN model Layer Layer type Image Input 2-D convolution ReLU Cross-channel 2-D Max Pooling 2-D convolution ReLU Cross-channel Normalization 2-D Max Pooling 2-D convolution ReLU Cross-channel 2-D Max Pooling 2-D convolution ReLU Cross-channel 2-D Max Pooling 2-D convolution ReLU Cross-channel 2-D Max Pooling 2-D convolution ReLU Cross-channel 2-D Max Pooling Concatenation Batch Normalization ReLU Fully connected ReLU Dropout Fully Connected Softmax Properties Activation Learnable Number of 227y227y3 images with Aozero centerAo 227 (S)y227 (S)y227 (S)y3 (C)y1 (B) 32 3y3 convolutions with stride .
227 (S)y227 (S)y32 (C) xy1 (B) Weight: 3y3y3 and padding AosameAo Bias: 1y1y32 ReLU 227 (S)y227 (S)y32 (C)y1 (B) Cross-channel normalization with 5 227 (S)y227 (S)y32 (C)y1 (B) channels per element 5y5 max pooling with stride .
and 227 (S)y227 (S)y32 (C)y1 (B) padding AosameAo 32 3y3 convolutions with stride .
227 (S)y227 (S)y32 (C)y1 (B) Weight: 3y3y32 and padding AosameAo Bias: 1y1y32 ReLU 227 (S)y227 (S)y32 (C)y1 (B) Cross-channel normalization with 5 227 (S)y227 (S)y32 (C)y1 (B) channels per element 5y5 max pooling with stride .
and 227 (S)y227 (S)y32 (C)y1 (B) padding AosameAo 32 3y3 convolutions with stride .
227 (S)y227 (S)y32 (C)y1 (B) Weight: 3y3y32 and padding AosameAo Bias: 1y1y32 ReLU 227 (S)y227 (S)y32 (C)y1 (B) Cross-channel normalization with 5 227 (S)y227 (S)y32 (C)y1 (B) channels per element 5y5 max pooling with stride .
and 227 (S)y227 (S)y32 (C)y1 (B) padding AosameAo 32 3y3 convolutions with stride .
227 (S)y227 (S)y32 (C)y1 (B) Weight: 3y3y32 and padding AosameAo Bias: 1y1y32 ReLU 227 (S)y227 (S)y32 (C)y1 (B) Cross-channel normalization with 5 227 (S)y227 (S)y32 (C)y1 (B) channels per element 5y5 max pooling with stride .
and 227 (S)y227 (S)y32 (C)y1 (B) padding AosameAo 32 3y3 convolutions with stride .
227 (S)y227 (S)y32 (C)y1 (B) Weight: 3y3y32 and padding AosameAo Bias: 1y1y32 ReLU 227 (S)y227 (S)y32 (C)y1 (B) Cross-channel normalization with 5 227 (S)y227 (S)y32 (C)y1 (B) channels per element 5y5 max pooling with stride .
and 227 (S)y22y(S)y32 (C)y1 (B) padding AosameAo 32 3y3 convolutions with stride .
227 (S)y227 (S)y32 (C)y1 (B) Weight: 3y3y32A and padding AosameAo Bias: 1y1y32 ReLU 227 (S)y227 (S)y32 (C)y1 (B) Cross-channel normalization with 5 227 (S)y227 (S)y32 (C)y1 (B) channels per element 5y5 max pooling with stride .
and 227 (S)y227 (S)y32 (C)y1 (B) padding AosameAo Concatenation of 2 inputs along the 227 (S)y454 (S)y32 (C)y1 (B) dimensional value Batch normalization 227 (S)y454 (S)y32 (C)y1 (B) Offset: 1y1y32 Scale: 1y1y32 ReLU 227 (S)y454 (S)y32 (C)y1 (B) 10 fully connected layers 1 (S)y1 (S)y10 (C)y1 (B) Weights 10y3297A Bias 10y1 ReLU 1 (S)y1 (S)y10 (C)y1 (B) 50% dropout 1 (S)y1 (S)y10 (C)y1 (B) 10 fully connected layers 1 (S)y1 (S)y10 (C)y1 (B) Weights 10y10 Bias 10y1 Softmax 1 (S)y1 (S)y10 (C)y1 (B) Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4692-4704
Int J Elec & Comp Eng
ISSN: 2088-8708
Performance evaluation After implementing the model, various performance metrics are calculated to assess its effectiveness in a binary classification task of identifying bone fractures in X-ray images.
The first step involves generating a confusion matrix, then key metrics such as true-positive rate, false-negative rate, false-positive rate, and true-negative rate are derived.
Following this, accuracy, precision.
F1 score, and error are computed to determine the optimal model for bone fracture identification.
Their equations Au.
Ay are given below:
ycNycycyce ycEycuycycnycycnycyce ycNycycyce ycAyceyciycaycycnycyce yaycaycaycycycaycayc = ycNycEycI = yaycAycI = yaycEycI = ycNycAycI = ycNycuycycayco ycuycu.
ycuyce ycycaycoycyycoyce ycNycycyce ycEycuycycnycycnycyce ycNycycyce ycEycuycycnycycnycyce yaycaycoycyce ycAyceyciycaycycnycyce yaycaycoycyce ycAyceyciycaycycnycyce ycNycycyce ycEycuycycnycycnycyce yaycaycoycyce ycAyceyciycaycycnycyce yaycaycoycyce ycEycuycycnycycnycyce yaycaycoycyce ycEycuycycnycycnycyce ycNycycyce ycAyceyciycaycycnycyce ycNycycyce ycAyceyciycaycycnycyce yaycaycoycyce ycEycuycycnycycnycyce ycNycycyce ycAyceyciycaycycnycyce ycEycyceycaycnycycnycuycu = ya1 ycIycaycuycyce = y 100% .
y 100% .
y 100% .
y 100% .
ycNycycyce ycEycuycycnycycnycyce ycNycycyce ycEycuycycnycycnycyce yaycaycoycyce ycEycuycycnycycnycyce 2yycEycyceycaycnycycnycuycu yycIyceycaycaycoyco ycEycyceycaycnycycnycuycu ycIyceycaycaycoyco yaycycycuyc ycIycaycyce = .
y 100% .
y 100% .
y 100% yaycaycoycyce ycAyceyciycaycycnycyce yaycaycoycyce ycEycuycycnycycnycyce ycNycuycycayco ycuycu.
ycuyce ycycaycoycyycoyceyc y 100% RESULTS AND DISCUSSION In this study, three deep learning models explore the detection of bone fractures from digital images.
Rigorous image preprocessing is applied to train and validate the models using a testing dataset.
To measure the performance of the developed models, several performance metrics are computed to find the optimal The three models are trained with 8,884 images and validated with 1,899 images.
During the implementation phase, some cases of ablation study are carried out to determine the best-suited parameters for the proposed model.
The models are implemented using 100 epochs and a batch size of 64.
However, in the case of the traditional CNN and the pretrained MobileNet-V2 model, performance is not satisfactory, and issues arise during the implementation phase.
Consequently, a newly developed proposed PDCNN model performs very well.
Four ablation case studies are considered to validate the proposed model for bone fracture identification.
The performance of these studies is detailed below:
Performance of CNN and MobileNet-V2 In terms of performance, both the traditional CNN and MobileNet-V2 models exhibit fluctuating results, indicating underfitting issues.
The confusion matrices of the CNN and MobileNet-V2 models are presented in Table 3 and performance metrices shwon in Table 4.
From Table 4.
CNN provide 74.
accuracy while MobileNet-V2 gained 81.
20% accuracy.
Examination of Figure 6.
and Figure 7.
reveals accuracy (Figure 6.
and Figure 7.
and loss curves (Figure 6.
and Figure 7.
that further emphasize the underfitting problems experienced by both models CNN and MovileNet-V2.
Underfitting occurs when a model is unable to capture the underlying patterns in the data, resulting in poor performance and low accuracy.
In the case of the traditional CNN and MobileNet-V2 models, this underfitting phenomenon is evident in their inability to adequately learn from the training data, leading to inconsistent and suboptimal results.
Table 3.
Confusion matrix for applied three models Model
CNN
MobileNet-V2
PDCNN
Bone-Net: a parallel deep convolutional neural network-based bone A (Md.
Hasan Imam Bijo.
A ISSN: 2088-8708 Table 4.
Performance metrices for CNN and MobileNet-V2 model Model Accuracy CNN MobileNet-V2 81.
TPR
FNR
FPR
TNR Precision .
F1 Score Error rate .
Figure 6.
The .
accuracy graph and .
loss graph for traditional CNN model .
Figure 7.
The .
accuracy graph and .
loss graph for MobileNet-V2 model Performance of proposed PDCCN model with ablation study In the preceding section, both models failed to deliver satisfactory results.
Now, it's imperative to validate the proposed model to demonstrate that the newly developed PDCNN model serves as the optimal solution for bone fracture detection.
In the validation process, four detailed case studies are conducted by altering hyperparameters such as kernel size, loss function, pooling layer, and optimizer.
The results of the ablation study are presented in Table 5.
This comprehensive analysis aims to ascertain the effectiveness of the proposed model in detecting bone fractures and to identify the optimal configuration for achieving the highest performance.
From the ablation studies in Table 5, various convolutional layer kernel sizes have been Four kernel sizes .
, 3, 4, and .
are compiled and evaluated.
Notably, a kernel size of 3 achieved the maximum accuracy, reaching 88.
61%, with a relatively low per-epoch training time of 127 seconds.
As a result, a kernel size of 3 is selected for implementation in the PDCNN model.
To optimize performance, different loss functions are evaluated, including binary cross-entropy, categorical cross-entropy, and mean squared error.
The model attained its highest test accuracy of 89.
65% when utilizing the categorical cross-entropy loss function.
Consequently, this loss function is chosen for integration into the final Further experimentation involved comparing max-pooling and average pooling layers.
It is found that the model achieved its peak performance with the max-pooling layer, resulting in an accuracy of 90.
Moreover, the optimization process included testing four distinct optimizersAiSGD.
Adam.
RMSprop, and Int J Elec & Comp Eng.
Vol.
No.
October 2025: 4692-4704
Int J Elec & Comp Eng
ISSN: 2088-8708
AdamaxAieach with a learning rate of 0.
001 and a batch size of 64.
Remarkably, the highest test accuracy of 89% is achieved by adam optimizer, surpassing all previous tuning efforts.
After meticulous parameter tuning and selection, the final model showcased satisfactory Subsequently, fine-tuning parameters are considered to refine the PDCNN model, and the performance metrics are presented in Table 6.
Figure 8 illustrates the accuracy and loss curves for the PDCNN model, demonstrating a lack of overfitting or underfitting during training, see Figure 8.
and Figure 8.
Both the training and validation curves smoothly converge, with minimal disparity between them.
Furthermore, the loss curves steadily decrease from the initial to the final epoch, maintaining a small gap This stability in the convergence of accuracy and loss curves underscores the robustness and effectiveness of the developed PDCNN model in bone fracture detection.
Table 5.
Performance results of four case of abalation studies Ablation Study Changing kernel size Configuration Parameter EpochyTime Test Accuracy Finding 100y127 s Improved 100y127 s Improved 100y152 s Droped 100y163 s Droped Changing the loss function Binary cross-entropy loss 100y124 s Droped Categorical cross-entropy 100y124 s Improved Mean squared error 100y124 s Droped Changing pooling layer Max 100y124 s Improved Average 100y124 s Droped Changing optimizer SGD 100y124 s Drop Adam 100y124 s Highest RMSprop 100y124 s Droped Adamax 100y124 s Droped Table 6.
Performance metrices for proposed PDCNN model Model Accuracy TPR FNR FPR TNR Precision F1 Score Error rate PDCNN 92.
Figure 8.
The .
accuracy graph and .
loss graph for proposed PDCNN model Comperative analysis and discussions In comparison with previous studies presented in Table 7, our study surpasses the performance achieved by earlier works in bone fracture recognition.
Notably, our proposed PDCNN model outperforms all previously reported methods.
This remarkable achievement underscores the effectiveness and advancement of our approach in accurately detecting bone fractures.
To maintain methodological coherence, we incorporated the canny edge detection image preprocessing technique alongside augmented images.
Additionally, we conducted rigorous validation of the proposed model through ablation studies.
These efforts collectively contributed to significant improvements in accuracy ranging from 1.
89% to 17.
Such substantial enhancements highlight the success of our study in pushing the boundaries of bone fracture detection capabilities.
Bone-Net: a parallel deep convolutional neural network-based bone A (Md.
Hasan Imam Bijo.
A ISSN: 2088-8708 Table 7.
Comparision table of earlier work on bone fracture recognition Studies .
Dataset
This Study 10,783 NM: Not mentioned N/A: Not application Preprocessing Technique Augmentation Ablation Study Model
Accuracy Performance Gain
Canny edge detection SNAKE
Sobel edge detector
Meta Classifier Sobel edge detector
SVM
discrete wavelet transforms CNN
WFP
M3 filter
SIFT BPNN
Canny edge detection and BPNN
Conservative Filtering HTM
Canny edge detection SVM
N/A
Canny edge detection PDCNN
N/A
CONCLUSION
This research marks a significant advancement in the field of medical image analysis by evaluating the effectiveness of various CNN architectures in detecting bone fractures.
Our study highlights the crucial role of customized neural network designs specifically tailored to meet the complex requirements of medical imaging tasks.
Among the models tested, the PDCNN stood out, demonstrating superior performance compared to MobileNetV2 and traditional CNNs.
Through rigorous validation, including four ablation studies.
PDCNN consistently achieved higher accuracy, precision, and F1-Score.
This modelAos innovative architectural enhancements were pivotal in addressing the intricate details of bone fracture detection, resulting in its exceptional ability to clearly differentiate between fractured and non-fractured bones.
Preprocessing techniques such as segmentation, edge detection, and feature extraction further enhanced the modelAos diagnostic capabilities.
The success of PDCNN suggests that it can significantly improve the accuracy and reliability of bone fracture diagnoses, which is a notable milestone in medical research.
This modelAos simplicity combined with its high accuracy promises substantial advancements in current diagnostic practices, potentially transforming the way bone fractures are identified from X-ray images.
The practical implementation of PDCNN could streamline the diagnostic process, reducing reliance on manual identification and thereby saving time and resources in clinical settings.
Moreover, the underperformance of traditional CNN and MobileNetV2 models, which exhibited underfitting issues, indicates that further research and refinement are necessary to optimize neural network architectures for medical applications.
Our findings underscore the importance of developing and employing customized neural networks to enhance diagnostic accuracy.
This research highlights the potential for these advancements to improve healthcare outcomes by enabling more precise and efficient diagnosis of bone fractures.
Overall, this study not only contributes to a deeper understanding of CNN applications in medical imaging but also paves the way for future innovations that could revolutionize medical diagnostics.
REFERENCES