Advancements in Orthopaedic Arm Segmentation: A Comprehensive
Review
Abhishek Swami, Snehal Farande,
Atharv Patil, Atharva Parle, Vivekanand Mane, Prathamesh Thorat
Department of Computer Science & Engineering (AIML)
D. Y. Patil College of Engineering and Technology, Kolhapur
—————————————————————————————————
Abstract — The most recent advances in medical imaging that have transformed diagnosis, especially
in the case of interpreting X-ray images, are actively involved in the healthcare sector. The advent of
digital image processing technology and the implementation of deep learning models such as
Convolutional Neural Networks (CNNs) have made the analysis of X-rays much more accurate and
efficient. In this article, some essential techniques such as edge detection, region-growing technique,
and thresholding approach, and the deep learning models such as variants of YOLOv8 – which is the
best object detection and segmentation framework – are reviewed. We further investigate that the
traditional image processing techniques like segmentation are very much simple and provides the
alternative to the advanced methods as well. Our review gives useful knowledge on the practical usage
of the innovative and traditional approaches of manual X-ray interpretation. The discovered
information will help professionals and researchers to gain more profound knowledge in digital
interpretation techniques in medical imaging.
Keywords — Image segmentation, edge detection, region growing, thresholding, Canny Operator,
YOLO, mAP
INTRODUCTION
The introduction of image segmentation has been a huge game changer between computer vision
and machine learning and as much in the field of X-ray screenings in medical diagnosis. This
technique has emerged as irreplaceable when making a diagnosis and planning the treatment [1].
This modern-day technology may be a real game-changer in terms of administering medical
decision-making procedures with more precision and accuracy. With its pixel-by-pixel feature,
image segmentation is capable of isolating structures in radiological images, such as bones and
tissues with little or no anomalies. Thus, image segmentation is considered as a highly efficient
tool with a number of applications. The members of medical cared observed the bone mending
prowess in the detection of abnormalities, bone fracture and disease diagnosis.
1
Traditional approaches for medical image segmentation utilised by professionals or by hand-
crafted designs based on image processing operations have been thresholding [2], edge detection
[3], and morphological operations. These ways enable to a degree of the interpretability and
control, which are useful in some applications. The classic approaches to the segmentation task,
however, have obvious limitations because the images and the segmentation situations in medical
imaging are complicated and diverse. Customised algorithms can not keep up with processing
speed and low accuracy when it comes to numerous scans. In addition to this, the manual feature
extraction from medical images requires experts who have the expertise and experience. So, their
task is prone to humans' uncertainties.
In recent years, deep learning algorithms have been widely used in medical image segmentation
to overcome these concerns. Deep feature learning allows models to extract semantic information
from images, improving segmentation accuracy and allowing them to adapt to different medical
image datasets and tasks. Since convolutional neural network (CNN)-based segmentation models
have achieved remarkable results, the combination of image segmentation and X-ray diagnosis
has evolved in such a way that it not only increases the precision of medical evaluations but also
equips medical professionals to design more informed and clever treatment plans, improving
patient outcomes. However, despite these advances, there are still difficulties associated with
interpreting X-ray images for various diseases, considering the variations in patient anatomy, and
understanding different bone structures. Despite these obstacles, image segmentation has the
revolutionary potential to redefine medical diagnosis and healthcare decision-making. The basic
idea of automating the X-ray diagnosis procedure highlights the path towards a more
sophisticated, efficient, and significant medical future.
We will review the method for segmenting Regions of Interest (ROIs) in X-ray Images, primarily
using the YOLOv8 model [4]. We aim to take advantage of the strengths of Convolutional Neural
Network (CNN) architectures, specifically YOLOv8, to achieve precise and efficient
segmentation. The YOLOv8 model, known for its exceptional object detection capabilities, is the
foundation of our method. It allows us to precisely identify and localise ROIs across a wide range
of medical imaging, improving the accuracy and efficiency of the segmentation process [5].
2
OVERVIEW OF TRADITIONAL IMAGE SEGMENTATION TECHNIQUES
Traditional image processing techniques for segmentation are essential techniques used to
identify objects of interest within images [6]. Following are some traditional segmentation
techniques-
1. Thresholding: It is a simple image segmentation method where pixels in an image are
categorized into two classes based on whether their intensity values are above or below a
specified threshold. Mathematically, it can be formulated as follows: Let I(x, y) represent the
intensity value of a pixel at coordinates (x, y). The thresholding operation can be represented
as:
{0, otherwise
1, if I(x, y) > Tthreshold
T (x, y) =
(a) Original X-ray image (b) Image formed after thresholding
Figure 1. Result of segmentation using thresholding.
Hand X-ray image of size 1280 × 1280 pixels passed through thresholding model with threshold of 177
3
• Advantages:
i. Simple and computationally efficient.
ii. Easy to implement.
• Disadvantages:
i. Sensitive to noise and variations in pixel intensity.
ii. May require manual selection of threshold value, which can be subjective.
2. Region Growing: It is a method where similar adjacent pixels are grouped together to form a
region based on predefined criteria such as intensity or texture similarity [7]. Mathematically,
it can be formulated as follows: Let I(x, y) represent the intensity value of a pixel at
coordinates x and y in the image. Let R represent the region being grown, initially containing
the seed point S. The region growing process can be represented as an iterative algorithm:
(a) Start of Growing a Region (b) Growing process after few iterations
Figure 2. Working of Region Growing Algorithm
I. Initialize R with the seed point S as shown in Figure 2(a).
II. For each pixel P in the region R:
a. Find neighbouring pixels of P.
b. Calculate the similarity between the intensity of each neighbouring pixel and
the average intensity of the pixels in R.
c. If the intensity of a neighbouring pixel is similar to the average intensity of R
(based on a predefined threshold), add that pixel to R.
III. Repeat step 2 until no more pixels can be added to R or until a stopping criterion is
met.
4
• Advantages:
i. Can handle irregularly shaped objects and regions.
ii. Automatically adapts to local image characteristics.
• Disadvantages:
i. Sensitive to seed point selection.
ii. Computationally expensive for large images or complex scenes.
(a) Original X-ray image (b) Image formed using region growing technique
Figure 3. Result of segmentation using region growing model
Hand X-ray image of size 1280 × 1280 pixels passed through region growing model at seed point (640,790)
with threshold of 60
3. Edge-based Detection: There are several methods in edge-based detection which aim to
identify edges or boundaries in an image by detecting discontinuities in intensity or texture.
Various edge detection operators can be used, such as Sobel, Prewitt, Roberts, and Canny.
These operators typically involve convolving the image with a kernel to compute gradients or
edge strengths.
3.1. Sobel Operator: It is commonly used for edge detection. It calculates the gradient of
the image intensity at each pixel, emphasizing edges where there is a significant
change in intensity [8]. It consists of two 3x3 convolution kernels, one for horizontal
changes and the other for vertical changes:
5
−1 0 1 −1 −2 −1
[−1 0 1] [1 1]
Gx = −2 0 2 Gy = 0 0 0
2
(a) Original X-ray image (b) Image formed after applying Sobel
Operator
Figure 4. Result of Edge-based Detection using Sobel Operator
Hand X-ray image of size 1280 × 1280 pixels operated with Sobel operator (kernel) of size 3 × 3
3.2. Prewitt Operator: It is similar to Sobel operator and is used for edge detection. It
calculates the gradient of the image intensity at each pixel in the horizontal and
vertical directions [9]. It consists of two 3x3 convolution kernels:
−1 0 1 −1 −1 −1
[−1 0 1] [1 1]
Gx = −1 0 1 Gy = 0 0 0
1
6
(a) Original X-ray image (b) Edge mask formed after applying Prewitt
operator
Figure 5. Result of Edge-based Detection using Prewitt Operator
Hand X-ray image of size 1280 × 1280 pixels operated with Prewitt operator (kernel) of size 3 × 3
3.3. Roberts Operator: It is a simple edge detection operator that calculates the gradient of
the image intensity using 2x2 convolution kernels [10]. It consists of two 2x2
convolution kernels:
[0 −1] [−1 0]
1 0 0 1
Gx = Gy =
7
(a) Original X-ray image (b) Edge mask formed after applying Roberts
Operator
Figure 6. Result of Edge-based Detection using Roberts Operator
Hand X-ray image of size 1280 × 1280 pixels operated with Roberts operator (kernel) of size 2 × 2
3.4. Canny Operator: It is a multi-stage algorithm used for edge detection, which includes
smoothing, gradient calculation, non-maximum suppression, and edge tracking by
hysteresis [11]. The specific kernel matrices used in Canny operator are not fixed as
they depend on the Gaussian smoothing and gradient calculation steps, which are
usually performed using convolution. These are the matrices for some commonly
used edge detection operators. These operators are applied to the image using
convolution to detect edges or boundaries.
• Advantages:
i. Can provide precise delineation of object boundaries.
ii. Robust to noise.
• Disadvantages:
i. May produce multiple edges in complex scenes.
ii. Prone to variations in intensity and noise.
8
(a) Original X-ray image (b) Edge detected using Canny Operator
Figure 7. Result of Edge-based Detection using Canny Operator
4. Active Contour Model (Snake): They are deformable models that evolve to fit the boundaries
of objects in an image based on energy minimization principles. Active contour models are
represented by a curve C(s) , where s represents the arc-length parameter [12]. The energy
functional E(C ) is minimized iteratively to deform the contour towards object boundaries.
(a) Original X-ray image (b) Edge detection using active contour model
Figure 8. Result of Edge-based Detection using Active Contour Model (Snakes)
9
• Advantages:
i. Can handle complex shapes and irregular boundaries.
ii. Robust to noise and occlusions.
• Disadvantages:
i. Sensitive to initialization and parameter tuning.
ii. Computationally intensive, especially for large images or complex objects.
These are basic explanations of each segmentation method along with their mathematical
formulations, results, advantages, and disadvantages. Further details and optimizations can be
explored based on specific applications and requirements. Thresholding stands out as a basic yet
efficient technique among these. It differentiates objects from the background by applying an
intensity threshold, classifying pixels with intensities above the threshold as part of the object and
those with values below the threshold as background. Moreover, morphological treatments such
as erosion and dilation change the shape and structure of segmented objects [13].
These methods, which are frequently used for post-processing, can refine segmentation results
and improve object connectivity. While these classic methods have their advantages, they might
fail when faced with complicated, noisy, or highly changeable images. These methods have
significant constraints in the context of X-ray images, which may limit their usefulness. X-ray
scans frequently show different levels of contrast, intricate anatomical features, and the
possibility of abnormalities. Because the intensity differences between objects and background
can be insignificant, thresholding techniques that require pixel intensity values might fail to give
correct segmentations in such scenarios. Besides, complex shapes and boundaries, which are
often seen in X-ray images, may offer difficulties for edge detection systems. Additionally, the
sensitivity of X-rays to image noise might result in the inclusion of undesired artefacts,
decreasing the accuracy of the segmentation results. Traditional techniques may also include
manual parameter adjustments, making them impractical for the quick processing of huge
quantities of X-ray images.
10
YOLOv8 AND ITS SIGNIFICANCE
YOLOv8 is a state-of-the-art object detection and image segmentation model, representing the
latest iteration of the YOLO series, which stands for "You Only Look Once” [14]. It has achieved
significant performance in computer vision and deep learning.
Figure 9. YOLOv8 Architecture
Here is an overview of YOLOv8 architecture:
• Backbone Network: YOLOv8 typically uses a powerful backbone convolutional neural
network (CNN) such as DarkNet or ResNet as its base. This backbone network is responsible
for extracting features from the input image at multiple scales and levels of abstraction.
• Feature Pyramid Network (FPN): YOLOv8 often incorporates a feature pyramid network
(FPN) or similar architecture to capture multi-scale features from different levels of the
backbone network. FPN helps in detecting objects of varying sizes and scales by combining
features from different layers.
11
• Detection Head: YOLOv8 employs a detection head or prediction module that predicts
bounding boxes, objectness scores, and class probabilities for the objects present in the
image. The detection head typically consists of a series of convolutional and fully connected
layers that process the features extracted by the backbone network.
• Anchor Boxes: YOLOv8 uses anchor boxes or priors to predict bounding boxes for objects.
These anchor boxes are predefined shapes with different aspect ratios and scales, which are
used as references for predicting the bounding box coordinates.
• Output Format: YOLOv8 outputs detections in a grid-based format. Each grid cell predicts
multiple bounding boxes along with confidence scores for object presence and class
probabilities. The final detections are obtained by applying non-maximum suppression
(NMS) to filter out redundant and overlapping bounding boxes.
• Loss Function: YOLOv8 typically uses a combination of loss functions, including localization
loss, confidence loss, and classification loss, to train the network. These loss functions
penalize errors in bounding box localization, objectness prediction, and class prediction,
respectively.
Significance:
• Real-Time Object Detection: YOLOv8 is known for its real-time object detection
capabilities, allowing it to recognise and locate objects within images or video frames
efficiently. This real-time capability is necessary for applications like surveillance,
autonomous vehicles, and augmented reality, where fast decision-making based on object
detection is essential [15].
• Single Forward Pass: YOLO's "You Only Look Once" concept tells that it performs object
detection and localisation in a single forward pass through the neural network. This
efficiency is the opposite of other object detection methods that require multiple passes,
making YOLOv8 faster and more computationally efficient.
12
• Versatility: Versatility is a significant aspect of YOLOv8, making it suitable for various
object detection tasks, including detecting objects in natural scenes, tracking objects over
time, and even segmenting objects.
• Accurate Object Localisation: YOLOv8 provides precise object localisation, meaning it
not only detects objects but also accurately outlines their boundaries. This high degree of
accuracy is crucial in applications where knowing the exact object location is essential,
such as in medical image analysis or autonomous navigation.
• Open Source and Community Support: YOLOv8, like its older versions, is open-source,
making it accessible to the researchers and developer community. This open nature
encourages collaboration, improvement, and the development of diverse applications.
• Transfer Learning: YOLOv8 can be fine-tuned and trained for specific tasks through
transfer learning, which enables researchers and developers to build custom object
detection models with less data and effort.
• State-of-the-Art Performance: YOLOv8 has pushed the boundaries of object detection
performance. It consistently achieves high mean average precision (mAP) scores on
benchmark datasets, indicating its state-of-the-art accuracy.
• Practical Applications: YOLOv8's combination of speed and accuracy has led to its use in
various practical applications, including security and surveillance, self-driving cars,
robotics, medical image analysis, and more. It has the potential to improve the efficiency
and safety of various industries.
YOLOv8's significance lies in its ability to provide real-time, accurate, and versatile object
detection capabilities, making it a valuable tool for multiple applications in computer vision and
beyond. Its efficiency and open-source nature have made it a popular choice for researchers and
developers.
13
COMPARISON OF YOLOv8 VARIANTS
Real-time object detection has attracted a great deal of interest in the development of the YOLO
(You Only Look Once) series that balances precision and swiftness. It is the latest instalment
from which innovations have been made to offer adaptability for multiple object detection tasks
over its predecessors [16].
Figure 10. YOLO models comparison plots
This analysis compares different YOLOv8 variations based on performance metrics with a focus
on how effective they are in segmentation tasks. The comparison includes model size, mAP for
bounding boxes and masks, inference speeds on different hardware platforms as well as model
complexity measured by parameters and FLOPs. Thus, it is possible to understand what kind of
trade-off there is between the size of the model, accuracy and computational efficiency; this can
be helpful to researchers or practitioners in choosing an appropriate version for their specific
tasks.
Table 1 - Comparison of YOLO-segmentation models of different sizes
Speed A100
size mAPbox MAP mask Speed CPU
Model TensorRT params (M) FLOPS (B)
(pixels) (ms) (ms) ONNX (ms) (ms)
YOLOv8n-seg 640 36.7 30.5 96.1 1.21 3.4 12.6
YOLOv8s-seg 640 44.6 36.8 155.7 1.47 11.8 42.6
YOLOv8m-seg 640 49.9 40.8 317 2.18 27.3 110.2
YOLOv8I-seg 640 52.3 42.6 572.4 2.79 46 220.5
YOLOv8x-seg 640 53.4 43.4 712.1 4.02 71.8 344.1
14
The presented data provides a complete comparison of YOLOv8 versions optimized for
segmentation tasks. Distinctions between each variant make use of its model size in the form of
pixels counting the width and height of input images. The mAP scores concerning bounding
boxes as well as masks capture how well the detection of objects is done and their segmentation
respectively with higher values indicating more efficient performances. Furthermore, times taken
to run inference on CPU and GPU that are ONNX runtime and A100 TensorRT were analyzed to
show the real-time capabilities that these models have. Meanwhile, parameters count and FLOPs
provide insights into computational complexity for each version–this helps to guide choice-
making based on computation resources available as well as application requirements.
APPLICATION IN X-RAY IMAGE ANALYSIS
We are focusing on the analysis of X-ray images of arm bones, specifically the radius, ulna,
carpals, metacarpals and phalanges. This targeted study starts with the collection and structuring
of a specialized dataset comprising X-ray images of these specific arm bone structures. Making
accurate masks and labels to highlight and identify different bones in these photos is an essential
component of this process [17]. By generating labels in a format suitable for YOLOv8, which is
the “.txt” file format, we aim to enable the model to recognize and distinguish these bones with
accuracy, simplifying the diagnosis of X-ray images.
1. Dataset Preparation:
• Data Collection: Collecting X-ray images that feature arm bones. These images should
represent various arm bone structures, such as the radius, ulna, carpals, metacarpals and
phalanges.
• Data Organization: Structuring dataset in a way recommended for training YOLOv8 model
with enhanced and resized X-ray images.
2. Mask Creation and Bone annotation:
• Manual Annotation: Skilled annotators manually create masks by outlining and labelling each
specific arm bone part within the X-ray images, using specific annotator tool.
15
(a) An X-ray image of a hand annotated with (b) An X-ray image of a hand with layered mask
instance labels using a polygon tool
Figure 11. Mask Creation and Bone Annotation
• Bone Identification: Each segmented bone, such as the radius or ulna, is accurately labelled
with its name to denote its specific identity.
3. Label generation in YOLOv8 format:
• Label File Generation: This involves generating YOLOv8-compatible labels for bones. These
labels contain information about the class index, object coordinates, and dimensions.
With these steps, a dataset focused on X-ray images of bones, specifically structured for YOLOv8
segmentation can be prepared. This streamlined approach showcases the analysis of X-ray images
with higher accuracy, which can be helpful for the precise diagnosis of underlying conditions.
16
RESULT AND ANALYSIS
We assessed how well the segmentation model can identify and outline different parts of bones.
To do this, we used facility in YOLOv8 base model provided by ultralytics and the
Weights&Biases tool (WandB.ai) for the evaluation. The experimental work involved multiple
image preprocessing steps, the design of a custom CNN for bone fracture prediction, and the use
of YOLOv8 for instance segmentation of bone subparts which helped in the development of
RadianceAI, contributing to its enhanced X-ray diagnosis capabilities using machine learning.
• Model Performance
The YOLOv8x-seg model with 295 layers, 71726434 parameters, 0 gradients, 343.7 GFLOPs
was trained on custom data of X-Ray images of bones such as radius, ulna, carpels, metacarpals
and phalanges. Training of this model was carried out on AI server facility provided by college,
which consists of 6 NVIDIA RTX A5000 GPUs each with 24256 MB GPU memory. The training
process was carried out for 200 epochs, with the best performance achieved on the 93rd epoch.
After that, the model did not show significant improvement for 20 epochs and it encountered
early stopping.
Figure 12. YOLOv8 Segmentation model training epochs
The model was trained to detect and segment various bones and fractures in input arm X-
Ray image, and it successfully achieved this objective. The training time took over 1 hour and 8
minutes.
17
• Data Sources
For our analysis, we collected X-ray images of hands from a diverse range of data repositories
and platforms. These included well-known resources like Roboflow and Hugging Face, as well as
various government digital libraries.
Figure 13. Raw X-ray Images
18
• Evaluation Metrics
- Precision (P): This represents the proportion of correctly identified objects out of all the
objects the model identified as that class. A high precision means the model makes few
mistakes (low false positives).
- Recall (R): This represents the proportion of correctly identified objects out of all the actual
objects present in the image. A high recall means the model misses few objects (low false
negatives).
- mAP (mean Average Precision): This is a single value that summarises the overall
performance across different confidence thresholds. It essentially takes the average of the
Precision-Recall Curve (PRC) area for various thresholds. There are two variations
mentioned here:
- mAP50: This refers to the mAP calculated using a specific confidence threshold of 50%.
- mAP50-95: This refers to the mAP calculated by averaging the PR curve area across
multiple confidence thresholds ranging from 50% to 95%.
The performance of the model based on evaluation set was evaluated using various
metrics provided in following table -
Box Mask
Class Images Instances
Precision Recall mAP50 mAP50-95 Precision Recall mAP50 mAP50-95
all 37 759 0.88 0.824 0.898 0.674 0.876 0.821 0.891 0.617
carpal 37 204 0.862 0.673 0.865 0.581 0.85 0.665 0.837 0.472
fracture 37 10 0.715 0.6 0.719 0.315 0.715 0.6 0.719 0.403
metacarpal 37 163 0.905 0.94 0.96 0.801 0.905 0.94 0.96 0.718
phalanx 37 318 0.885 0.872 0.907 0.713 0.872 0.86 0.894 0.543
radius 37 31 0.946 0.935 0.976 0.817 0.945 0.935 0.976 0.808
ulna 37 33 0.968 0.925 0.96 0.814 0.968 0.925 0.96 0.759
Table 2 - Evaluation of model based on different metrics
19
• Interpreting the Table:
- Class: This indicates the category of object the model is trying to detect (e.g., carpal,
fracture).
- Images: This shows the number of images used for evaluation in this class.
- Instances: This represents the total number of actual objects present in the images for this
class.
- Box (Metric): These metrics refer to the performance of the model in detecting bounding
boxes around the objects.
- A high value (>0.8) in Box (P) and Box (R) indicates good precision and recall for
bounding boxes in that class.
- mAP50 and mAP50-95 for boxes provide a more comprehensive picture of the model's
performance across different confidence levels.
- Mask (Metric): Similar to bounding boxes, these metrics evaluate the model's performance
in predicting segmentation masks around the objects (if applicable).
Looking at the "carpal" class, we see Box (P) is 0.862, indicating the model is good at
identifying correct bounding boxes for carpals with relatively few false positives. However, Box
(R) is 0.673, suggesting the model might miss some actual carpals (false negatives). The mAP
values provide a more complete picture of performance across various confidence levels.
This table provides insights into how well the model performs in detecting different object
classes in the dataset. By analyzing precision, recall, and mAP, we can assess the trade-off
between correctly identifying objects and missing some.
20
• Confusion Matrix :
The confusion matrix is a table that summarises the performance of the model by showing
true vs predicted results across different bone classes. It helps in understanding how well the
model is classifying the objects.
As we can observe and analyse confusion matrix in Figure.14, True Positive Rate of our
model is higher since it is correctly identifying and and predicting mask for segmentation for
each bone type. We can also interpret that model can be trained on more data to reduce True
Negatives and False Positives.
Figure 14. Confusion Matrix
21
- F1 Confidence Curve :
The F1 confidence curve shows the relationship between the F1 score and the confidence
threshold. It helps in determining the optimal confidence threshold for the model. As it can be
observed from Figure 7.4 , F1-score increases as confidence threshold is increased for all classes
and suddenly drops after 0.963.
Figure 15. F1 vs Confidence curve
- Precision-Confidence Curve :
The precision curve shows the relationship between precision and confidence. In Figure.
7.5, it can be observed that, with confidence threshold, precision of the model increases. From
0.79 to 1.00 confidence threshold, an exponential increase in precision value can be observed.
Figure 16. Precision vs Confidence curve
22
- Precision-Recall Curve :
The precision-recall curve is a graph that shows the relationship between precision and
recall at different confidence thresholds. It helps in determining the optimal confidence threshold
for the model. In Figure 7.6, it can be observed that precision and recall are reciprocal of each
other at mean average precision i.e. mAP 0.5 for all classes.
Figure 17. Precision vs Recall curve
- Recall-Confidence Curve :
Recall confidence is a term commonly used in object detection models, such as YOLO
(You Only Look Once). It is a measure of how well the model can detect all the objects in an
image, without missing any. Figure 7.7 suggest that, confidence value of 0.6 can deliver better
results at optimal recall value.
Figure 18. Recall vs Confidence curve
23
• Validation Batch Outputs :
The outputs of training and validation batches were observed to analyse the model's performance
and make necessary adjustments. During the training process, the model's outputs were compared
with the ground truth labels to compute the loss function. The loss function was used to update
the model's parameters to improve its performance. The validation set was used to evaluate the
model's performance on unseen data.
Figure 19. Validation Batch Output
The YOLOv8-seg model trained on the given dataset has shown good performance in detecting
and segmenting different bones. The use of diverse images and the evaluation of various metrics
helped to assess the model's performance. However, further testing and validation on a larger
dataset with varying conditions may be required to confirm its robustness.
24
• Inference
Throughout our work with hand X-ray pictures, we've made a number of essential image
enhancements. This includes sharpening the photos to highlight critical features and adjusting
brightness and contrast with gamma correction to provide a sharper view of bone architecture.
Figure 20. X-ray images of a hand with predicted bone instance segmentation masks and boxes
Edge detection was also used in our investigation, which allowed us to enhance the curves and
edges of the hand's bones, making them stand out clearly. We next moved on to binarization,
which included converting the photos into a binary representation and separating the foreground
(which included the delicate hand anatomy) from the backdrop. These strategic image processing
procedures completed the path for accurate segmentation and extensive analysis, eventually
advancing medical diagnostics and research.
25
CHALLENGES AND LIMITATIONS
YOLOv8 has shown a state-of-the-art results in object detection and segmentation tasks, but as
we consider medical image analysis, we encounter distinct challenges and limitations. Challenges
include the intricacy of manual annotation, the variability in human anatomy, noise and artefacts
in X-ray images, and the limited data. Also, it operates within certain limitations, such as the
model's capacity to generalise, availability of enough computational resources etc.
• Challenges:
1. Precise annotation: Accurate manual annotations of various bones can be difficult due to
complex bone structures, which may overlap in X-ray images.
2. Anatomy variability: Human anatomy changes from person to person. Building a model that
generalises across various individuals might be a difficult task.
3. Noise and artefacts: X-ray scans often contain noise and artefacts, which can make the
segmentation task difficult.
4. Data limitations: Obtaining X-ray images can be tough, especially when working with
specific types of bones. Limited data may affect the model’s ability to generalise effectively.
• Limitations:
1. Generalisation: The model may have difficulty generalising new, unseen bone structures,
leading to inaccurate detection and segmentation.
2. Dependency on High-Quality Images: The model accuracy depends on the quality of the X-
ray images. Images with poor quality, low resolution or excessive noise can negatively impact
the segmentation results.
3. Computational Resources: Training and deploying deep learning models like YOLOv8 can be
computationally intensive, requiring access to powerful hardware and substantial
computational resources.
4. Human Expertise: Manual annotation and preparation of accurate masks and labels require
expertise in radiology and image analysis. Acquiring such skilled annotators can be resource-
intensive.
26
CONCLUSION
In this paper, we have explored the significant role of YOLOv8, a cutting-edge object detection
and segmentation framework, in X-ray image analysis, particularly in the context of bone
segmentation. The application of deep learning models, especially Convolutional Neural
Networks (CNNs), has transformed the landscape of medical diagnostics, specifically in X-ray
interpretation, enabling medical professionals to make more informed and accurate decisions in
various scenarios. Traditional segmentation methods, while useful in many contexts, have
inherent limitations when dealing with the complexity and variability of medical images,
particularly in medical image segmentation tasks. Deep learning algorithms, with their ability to
extract semantic information and adapt to diverse datasets, have addressed these limitations
effectively. CNN-based models, such as YOLOv8, have played a significant role in advancing
medical image analysis, enhancing both precision and efficiency. The combination of image
segmentation and X-ray diagnosis has the potential to reshape the field of healthcare, offering
automated, reliable, and detailed insights into various medical conditions.
In our research, we have focused on X-ray images of arm bones, specifically the radius, ulna,
carpals, metacarpals, and phalanges. We have outlined the steps for dataset preparation, mask
creation, and label generation, highlighting the format suitable for YOLOv8. The structured
dataset can enable YOLOv8 to recognize and differentiate these bone structures accurately,
ultimately simplifying and improving the diagnosis of X-ray images in healthcare. However, we
must acknowledge existing challenges and limitations in this field. From precise manual
annotation to the variability in human anatomy and the dependence on high-quality images. It is
essential to recognize the need for skilled annotators, the requirement for substantial
computational resources, and the potential for limitations in model generalisation.
In conclusion, YOLOv8 represents an important step in the application of deep learning to X-ray
image analysis, offering a promising solution for accurate and efficient bone segmentation. While
challenges and limitations persist, ongoing research and advancements in the field of medical
image analysis, driven by models like YOLOv8, hold the potential to revolutionise healthcare
diagnostics, clearing the path for a more advanced, effective, and meaningful medical future.
27
REFERENCES
[1] Cheng, J.Z., Ni, D., Chou, Y.H., Qin, J., Tiu, C.M., Chang, Y.C., Huang, C.S., Shen, D.,
Chen, C.M.: Computer-aided diagnosis with deep learning architecture: Applications to
breast lesions in us images and pulmonary nodules in CT scans. Scientific Reports 6, 24454
(2016)
[2] Otsu, N.: A threshold selection method from grey-level histograms. IEEE transactions
on systems, man, and cybernetics 9(1), 62–66 (1979)
[3] Magnier, Baptiste: Edge detection: a review of dissimilarity evaluations and a proposed
normalized measure. Multimedia Tools & Applications (2017)
[4] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. YOLO by Ultralytics, Jan. 2023.
[5] V. Kawade, V. Naikwade, V. Bora and S. Chhabria, "A Comparative Analysis of Deep
Learning Models and Conventional Approaches for Osteoporosis Detection in Hip X-ray
Images," 2023 World Conference on Communication & Computing (WCONF), RAIPUR,
India, 2023, pp. 1-7, doi: 10.1109/WCONF58270.2023.10235129.
[6] Zaitouna, N. M., & Aqel, M. J. (2015). Survey on Image Segmentation Techniques.
Procedia Computer Science, 65, 797-806. ISSN 1877-0509. DOI: 10.1016/
j.procs.2015.09.027.
[7] Ikonomatakis, N. & Plataniotis, Konstantinos & Zervakis, Michalis & Venetsanopoulos,
A.N.. (1997). Region growing and region merging image segmentation. 299 - 302 vol.1.
10.1109/ICDSP.1997.628077.
[8] Sobel, I., & Feldman, G. (1968). A 3x3 Isotropic Gradient Operator for Image Processing.
Presented at the Stanford Artificial Project.
28
[9] N. Prakash, S. Asif Basha, S. Chowdhury, B. Reshmi, D. Kapila and S. Devi,
"Implementation of Image Segmentation with Prewitt Edge Detection using VLSI
Technique," 2022 International Conference on Innovative Computing, Intelligent
Communication and Smart Electrical Systems (ICSES), Chennai, India, 2022, pp. 1-6, doi:
10.1109/ICSES55317.2022.9914259.
[10] Acharjya, Pinaki & Bera, Mihir. (2021). Detection of edges in digital images using edge
detection operators. Computer Science & Engineering An International Journal. 9. 107-113.
[11] Canny, J. (1986). A Computational Approach to Edge Detection. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 8(6), 679-698.
[12] Rj, Hemalatha & Tr, Thamizhvani & Dhivya, A. & Joseph, Josline & Babu, Bincy &
Chandrasekaran, R.. (2018). Active Contour Based Segmentation Techniques for Medical
Image Analysis. 10.5772/intechopen.74576.
[13] Serra, J. (1983). Introduction to Mathematical Morphology. Computer Graphics and Image
Processing, 22(4), 385-393.
[14] Jacob Solawetz and Francesco.What is YOLOv8? The ultimate guide., 2023. 04-30-2023
[15] Juan R. Treven and Diana M. Cordova-Esparaza. A comprehensive review of yolo: From
yolov1 to yolov8 and beyond, 2023. Supplied as additional material
https://arxiv.org/pdf/2304.00501.pdf
[16] Jocher, G., Chaurasia, A., & Qiu, J. (2023). Ultralytics YOLOv8 (Version 8.0.0) [Software].
Retrieved from https://github.com/ultralytics/ultralytics
[17] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, Sanja Fidler : Annotating Object
Instances with a Polygon-RNN
https://www.cs.toronto.edu/~urtasun/publications/castrejon_etal_cvpr17.pdf
29
View publication stats
