A study conducted by the STEAM in AI: Data + AI Ethics Intensive Summer Cohort, providing valuable insights into the evolving intersection of AI and healthcare enhancing Knee MRI Analysis.
Research conducted by Julia Kessman, Yidian Song, Nicholas Vlasceanu, Siyona Agarwal and Alexa Torres under the guidance of Mentor Juliana Shihadeh, PhD Computer Engineering at Santa Clara University.
This research used data provided by the Stanford Center for Artificial Intelligence in Medicine and Imaging (AIMI). AIMI curated a publicly available imaging data repository containing clinical imaging and data from Stanford Health Care, the Stanford Children’s Hospital, the University Healthcare Alliance and Packard Children’s Health Alliance clinics provisioned for research use by the Stanford Medicine Research Data Repository (STARR).
Advancements in artificial intelligence (AI) have revolutionized data processing across various domains, with healthcare emerging as a prominent beneficiary of AI-driven innovations. This paper delves into the application of AI in healthcare, specifically focusing on improving the accuracy of diagnosing knee injuries, particularly anterior cruciate ligament (ACL) tears, through the analysis of MRI images. The research explores the utilization of the K-Means clustering algorithm for classifying knee MRI images and examines the potential of using RGB pixel values extracted from these images.
Additionally, the study investigates the impact of viewing knee images from various angles to enhance classification accuracy. The experimental results reveal that a combination of K-Means clustering and RGB pixel analysis, while promising, falls short of providing a highly accurate ACL tear classification. Graphical representations are employed for initial model evaluation, demonstrating the model’s ability to distinguish between images with and without ACL tears. However, it becomes apparent that more sophisticated AI models may be necessary when dealing with intricate data, such as identifying knee tears.
The paper concludes by emphasizing the need for further research, highlighting the potential of deep learning models in future investigations to enhance the accuracy of ACL tear classification. This research underscores the importance of leveraging advanced AI techniques to address complex medical diagnoses, thereby advancing the capabilities of AI in the field of healthcare. The study is conducted by the STEAM in AI Data + AI Ethics Summer Cohort, providing valuable insights into the evolving intersection of AI and healthcare.
The integration of artificial intelligence (AI) in the healthcare sector has seen remarkable growth in recent years. This surge is primarily driven by its potential to enhance diagnostic accuracy, streamline operational workflows, and simplify patient monitoring. In healthcare, computer-based technologies are harnessed to process digitized data from diverse sources, including computed tomography (CT), magnetic resonance imaging (MRI), X-ray, and ultrasound. Notably, deep learning has displayed exceptional performance, particularly in computer vision, when analyzing medical images and videos. The continual advancements in deep learning and image processing have enabled the detection and localization of intricate patterns across various radiological imaging modalities, with some achieving results on par with human decision-making.
Traditionally, radiologists meticulously examine entire medical images to identify anomalies, analyze their attributes, and subsequently document their observations. However, this process is time-consuming, and crucial abnormalities may occasionally evade human detection. Technological advancements have accelerated the generation of high-resolution MRI images, expediting the scanning process. Deep learning techniques have been extensively employed in diagnosing medical conditions through MRI data across different anatomical regions.
The objective of this project was to develop an effective AI model capable of classifying knee MRI images as either healthy or indicative of an anterior cruciate ligament (ACL) tear. This research introduces an AI classification model tailored for knee diagnostics, which processes MRI images to determine the presence of an ACL tear and overall knee health.
Through this project, we gained valuable insights into data handling, ethical considerations in AI, AI and machine learning methodologies, Python programming, and the development of a K-Means clustering model for image classification. The central focus of the project was the creation of a model that could analyze diverse MRI images to assess knee health.
This research is situated within the dynamic and evolving landscape of AI’s role in assisting healthcare professionals in making accurate diagnoses. We note a recent study that aligns with our project’s objectives—a scientific journal article explaining the use of deep learning to identify issues in MRI images, thus aiding medical professionals.
Our generation is actively seeking innovative ways to integrate technology into various aspects of life, including sports, academics, arts, and more. Recognizing that technology will play a pivotal role in our future and careers, we are committed to contributing to the development of machine learning and AI models that support medical professionals in their work.
Before embarking on any substantial work, it is imperative to provide an overview of the project and acknowledge the ethical concerns it raises. Ethical considerations, such as skewed data or code bugs introducing bias and errors into the model, must be addressed proactively through measures like rigorous code review and bias testing. Additionally, the ethical concern of misidentifying MRI images and potential life-threatening consequences necessitates a collaborative approach where AI augments rather than replaces human expertise.
Reviewing prior work for inspiration and insights is a crucial step in any project’s inception. Notably, scientists in Singapore have explored the use of AI to interpret MRI brain scans and predict mental states, even envisioning its future application in virtual reality (VR). However, this technology has faced challenges related to overfitting due to limited data, highlighting the importance of learning from both successes and failures in interpreting MRI images using AI.
Our dataset, obtained from the Stanford University Medical Center and titled MRNet, comprised MRI knee scan images captured from three different angles: axial, sagittal, and coronal. Each set of images presented distinct classifications related to knee health, including ACL tears, meniscal tears, and no tears. Our project focused solely on building a model to detect ACL tears. The MRNet dataset conveniently provided separate training and validation sets, facilitating our model development process.
For our project, each student analyzed 500 MRI knee scan images, focusing on a specific angle. At least one student examined each angle. The initial step involved segregating the training and validation data into distinct lists, with a split ratio of 80% for training and 20% for validation. Labels indicating the presence of an ACL tear or its absence were extracted from the dataset.
We adopted a K-Means clustering model for our project. K-Means is a clustering algorithm that identifies cluster centers based on mean values derived from input data. In our case, we employed two classes: ACL tear and no ACL tear. Initially, the model determined the means of pixel values from the images. In the subsequent training phase, the model iteratively adjusted mean values to minimize the sum of squared distances between data points and cluster centers. This phase aimed to refine the accuracy of cluster center selection. The training accuracy was computed to assess how well the model classified the training data. However, the validation accuracy, determined by testing the model on the reserved 100 images, provided a more comprehensive evaluation of its effectiveness.
A graphical representation of the K-Means clustering model was generated to illustrate the clustering of validation data post-training. In an ideal scenario, data points should form distinct clusters with no overlap, indicating accurate classification. However, in our case, significant overlap was observed between data points in the two clusters, indicating room for improvement in model accuracy.
We identified data imbalances as a significant issue in our project, requiring remediation. This could be addressed through undersampling or oversampling techniques, enhancing result accuracy. While we did examine accuracy briefly, we acknowledge that further in-depth analysis is warranted in future iterations. Deep learning emerged as a promising avenue for future exploration due to its capability to handle large, complex datasets and its “automatic learning” feature, reducing the need for hand-engineered features.
Furthermore, we can extend our project by incorporating additional data types, such as axial, coronal, and sagittal MRI images, and potentially incorporating information related to meniscal tears and other knee conditions. Experimenting with diverse data sources and types presents an exciting avenue for future research and development in this field.