Although attention in AI related to healthcare is expanding, there hast been much progress in translating or implementing these technologies for clinical usage. Therefore, as we conduct our research, we will open a new field of study for the integration of artificial intelligence (AI) in medical assessment tools and physical therapy field. This will save physiotherapists time and effort for developing Proper evaluations and conduction of treatment plans, benefit patients and the country overall by lowering the financial burden associated with making accurate evaluations and management for these cases, and provide clear, objective evaluations for the majority of spinal scoliosis deformities as well as appropriate personalization for treatment plans.
Martial and Methods
1. Experiment martials This section presents the findings of the research, detailing the methodologies employed and the outcomes obtained. The analysis begins with an overview of the dataset used, followed by the preprocessing steps undertaken to ensure data quality and suitability for model training. then delve into the various approaches and models selected for this study, including both custom Convolutional Neural Networks (CNNs) and pre-trained models utilized through transfer learning. Each model is assessed based on its performance, and discuss the experimental setup and evaluation metrics used to measure effectiveness. Finally, presenting the experimental results, providing a comprehensive analysis of model accuracy, precision, recall, and other relevant metrics.
2.1 Dataset collection The data of the subjects were divided for 2 groups. Group (A) normal spine total X-ray consisted of 664 image and group (B) scoliotic patients x-ray consisted of 4307 images. All spinal X-ray involved in this study were retrospectively compiled from manifold sources, including BUU Datasets , Kaggle, Mendeley Data, Huggingface, Dropbox , and Roboflow, ensuring a diverse and comprehensive collection of scoliosis-related images. Patient with scoliosis were (1) diagnosed with scoliosis for different etiology, (2) Clear X- Ray for spine for all spinal curvatures including; cervical, thoracic and lumber. (3) C shaped and double C shape scoliosis. (4) mild, moderate and sever scoliotic degrees. (5) Adolescents with mean age 17(6) Both gender(male and female).
2. Purposed methodology 2.1 Data Augmentation To address the issue of limited data, we employed data augmentation to artificially expand our dataset. This technique involves generating new training samples by applying various transformations to the existing images, such as flipping, scaling, and rotating. These augmentations not only increase the dataset size but also play a crucial role in enhancing the model ability to generalize by exposing it to a wider variety of image variations. By simulating different viewing conditions and distortions, data augmentation helps the model become more robust to changes in image orientation, scale, and other variations that it may encounter in real-world scenarios. This process is particularly important in medical image analysis, where acquiring large, diverse datasets can be challenging. Data augmentation ensures that the model does not overfit to the limited original dataset and instead learns to recognize the underlying patterns that are indicative of scoliosis, ultimately improving its performance and reliability.
Original images Flipped image Rotated image Scaled image
. Examples of Data Augmentation Techniques Applied Images 2.2 Preprocessing Effective preprocessing is a critical step in preparing our dataset for model training, as it enhances the quality and consistency of the images, making them more suitable for analysis. One of the key techniques was employed histogram equalization, which is particularly important in medical imaging where subtle variations in pixel intensity can significantly impact the accuracy of classification. Histogram equalization improves the contrast of images by redistributing their intensity values, ensuring that the histogram-or distribution of pixel intensities-is more evenly spread out. This process enhances important features in the images, making them more informative and easier for the model to interpret.
2.3 Approaches and Model selection This study investigating various machine learning and deep learning models to tackle the complex task of classifying Normal and Scoliosis images. The research approach was meticulously designed to strike an optimal balance between achieving high classification accuracy and maintaining computational efficiency, which is crucial for real-world medical applications. A careful selection for range of models, each offering unique strengths in handling image data, and configured them to maximize their performance on the inserted dataset. The following paragraphs provide a comprehensive overview of the models employed, the rationale behind their selection, their specific configurations, and a detailed analysis of the results obtained from the experiments. This thorough examination not only highlights the effectiveness of the chosen models but also offers insights into their potential applicability in clinical settings.
2.3.1 Convolutional Neural Networks (CNNs) utilized Convolutional Neural Networks (CNNs) for exceptional capability in image classification tasks. CNNs excel at capturing spatial hierarchies and features within images through a combination of convolutional layers, pooling layers, and dense layers Convolutional Layers: These layers are fundamental to CNNs, as they apply various filters to the input images to detect patterns such as edges, textures, and shapes. By sliding these filters across the image, convolutional layers generate feature maps that represent different aspects of the image. We experimented with various filter sizes and numbers of filters to optimize the feature extraction process.
Pooling Layers: Pooling layers, typically using operations such as max pooling or average pooling, down-sample the feature maps produced by convolutional layers. This reduces the spatial dimensions of the feature maps, thereby decreasing the computational load and helping the model generalize better by focusing on the most salient features. We employed pooling layers to effectively manage the dimensionality and improve the model efficiency.
Dense Layers: After the convolutional and pooling operations, dense layers (fully connected layers) used to interpret the features extracted by the convolutional layers. Dense layers aggregate the high-level features and perform the final classification by mapping the learned representations to the output labels. We tuned the number of dense layers and their units to enhance the model ability to make accurate predictions.
This combination of convolutional, pooling, and dense layers allowed us to build a robust CNN architecture that effectively distinguishes between Normal and Scoliosis images.
2.3.2 Transfer Learning
Given the limited size of our dataset, we leveraged transfer learning to enhance model performance without extensive computational resources. Transfer learning involves using pre-trained models-developed on large, diverse datasets-and adapting them to our specific task. several advanced pre-trained models to benefit from their learned features :
XceptionNet: XceptionNet is an architecture based on depth-wise separable convolutions, which decomposes convolution operations into smaller, more efficient operations. This design helps capture complex features while reducing computational complexity. We fine-tuned XceptionNet to adapt its learned features for our scoliosis classification task.
DenseNet201: DenseNet201 is characterized by its dense connectivity pattern, where each layer receives input from all preceding layers. This approach promotes feature reuse and improves gradient flow during training. DenseNet201 utilized to leverage its strong feature extraction capabilities and enhance our model accuracy.
EfficientNetB0: EfficientNetB0 employs a compound scaling method that uniformly scales the depth, width, and resolution of the network. This model is known for its balance between accuracy and efficiency. EfficientNetB0 adapted to the dataset to benefit from its optimized architecture and high performance.
InceptionV3: InceptionV3 features a complex architecture with multiple parallel convolutional operations of varying filter sizes, enabling it to capture a wide range of features. InceptionV3 incorporated to take advantage of its robust feature extraction and multi-scale processing abilities.
MobileNetV2: MobileNetV2 is designed for mobile and edge devices with efficiency and speed in mind. It uses depth-wise separable convolutions to reduce model size and computational requirements. MobileNetV2 used to achieve a balance between model efficiency and classification accuracy.
By fine-tuning these pre-trained models, makes the classifiers able to leverage their advanced feature extraction capabilities and adapt them to selected specific task of classifying Normal and Scoliosis images, resulting in improved performance and reduced training time.
