Detection of Knee Pathologies from MRI Using Deep Learning Models (CNNs) - Spring 2026

This is a Report Template Quarto

Author

Sabna Pillai (Advisor: Dr. Cohen)

Published

Invalid Date

Slides: slides.html ( Go to slides.qmd to edit)

Important

Remember: Your goal is to make your audience understand and care about your findings. By crafting a compelling story, you can effectively communicate the value of your data science project.

Carefully read this template since it has instructions and tips to writing!

Nice report!

Introduction

The incidence of musculoskeletal disorders, especially those involving the knee, including osteoarthritis (OA), anterior cruciate ligament (ACL) injuries, and meniscal injuries, is a major global healthcare concern that has a profound effect on the mobility and quality of life of affected patients. Conventional diagnostic tools, especially X-ray radiography, have several limitations, including the inability to image soft tissues, often resulting in the missed diagnosis of early degenerative changes until late-stage damage has been incurred (Panwar et al. 2025). Magnetic Resonance Imaging (MRI) has recently become the standard diagnostic tool for musculoskeletal imaging because of its high contrast resolution and multi-planar imaging capabilities, which enable the detailed evaluation of ligaments, cartilage, and bone marrow (Qiu et al. 2021). However, the current manual analysis of knee MRI scans is a time-consuming and error-prone process, especially with regard to inter-observer variability and the high cognitive load of radiologists, which can result in delays in diagnosis and treatment (Bien et al. 2018).

Conventional MRI assessment of the knee relies on manual segmentation of anatomical structures, which is time-consuming and prone to variability. Zhou et al. (2018) demonstrated that a deep convolutional neural network can automatically segment knee joint tissues, including cartilage and meniscus, providing accurate and consistent delineation to support more efficient analysis of knee anatomy and pathology (Zhou et al. 2018)

The synergy of artificial intelligence, particularly deep learning (DL) and Convolutional Neural Networks (CNNs), provides a revolutionary approach to overcome these diagnostic hurdles by enabling an automated, objective, and scalable analysis of complex medical images (Oettl et al. 2025). CNNs are especially suited for medical imaging applications because they have the capability to automatically learn hierarchical feature representations directly from raw pixel data, including subtle pathological patterns that may not be visible to the human eye (Yeoh et al. 2021). Recent breakthroughs have shown the effectiveness of both 2D and 3D CNN models for segmenting musculoskeletal tissues and grading the severity of structural damage (Liu et al. 2018; Guida, Zhang, and Shan 2021). However, there are still major challenges to overcome in the current state of research, such as the requirement for large and high-quality annotated datasets and the lack of generalizability of models across different MRI protocols and hardware configurations (Goel 2025).

To overcome these challenges, there is an urgent need for the development of efficient automated systems that are capable of multi-pathology detection and grading to help doctors in high-throughput settings (Astuto et al. 2021). Although specialized models have proved to be successful in specific tasks like the detection of ACL tears using self-supervised learning techniques (Aidarkhan et al. 2025) or staging meniscus degeneration (Pedoia et al. 2019), the development of a comprehensive framework that can distinguish between different simultaneous knee injuries is still an intricate challenge. Using the latest advancements in CNN models, this research aims to develop an automated system for the detection and grading of knee pathologies from MRI images, which can eventually help in providing better patient care strategies for early intervention (Awan et al. 2021).

Manual assessment of knee osteoarthritis severity is often time-consuming and subjective. A recent study demonstrated that a CNN-based deep learning model can automatically detect and grade knee osteoarthritis from imaging data, providing accurate and consistent classification that can support faster and more reliable diagnosis.(Rani et al. 2024)

This research work is driven by the need to enhance the accuracy and efficiency of diagnosis in orthopedic radiology.

Methods

Convolutional Neural Network (CNN)

A Convolutional Neural Network (CNN) is a specialized deep learning architecture designed to process structured, grid-like data such as images. CNNs excel in computer vision and medical imaging because they autonomously learn hierarchical spatial features from raw pixel inputs, significantly reducing the reliance on manual feature extraction.

In the initial convolutional layers, the network captures basic spatial features like intensity gradients and tissue boundaries. As the network deepens, it progressively learns more complex and abstract anatomical patterns, such as cartilage degeneration or disruption of anterior cruciate ligament (ACL) fibers. This hierarchical feature extraction allows CNNs to represent intricate anatomical structures effectively.

While 1D CNNs typically handle sequential data and 2D CNNs focus on individual image slices, 3D CNNs expand convolution operations into the volumetric domain by incorporating depth alongside height and width. This volumetric approach preserves spatial context across multiple slices, which is critical for imaging modalities like magnetic resonance imaging (MRI).

In knee MRI applications, 3D CNNs have shown enhanced accuracy in detecting and grading ACL tears, meniscus degeneration, and osteoarthritis severity. By maintaining inter-slice spatial continuity, these models provide a richer anatomical understanding, positioning 3D CNNs as a powerful tool in cutting-edge orthopedic artificial intelligence systems (Bien et al., 2018; Guida et al., 2021; Oettl et al., 2025).

Basic Architecture of a CNN

A typical CNN consists of the following components:

Basic Architecture of a CNN
  1. Convolution Layer : Applies filters (kernels) to extract features from input images.

    Mathematical representation: \[ Y(i,j) = \sum_{m}\sum_{n} \left[ X(i+m, j+n)\cdot W(m,n) \right] + b \]

    Where: X= input image W= filter/kernel b= bias Y= output feature map

  2. Activation Function:

    Commonly ReLU: \[ f(x) = \max(0, x) \]

  3. Pooling Layer Reduces spatial dimensions (e.g., Max Pooling).

  4. Fully Connected Layer Performs classification based on extracted features.

  5. Output Layer Uses Softmax (multi-class) or Sigmoid (binary/multi-label).

In medical imaging, CNNs have demonstrated strong performance in MRI-based diagnosis and tissue segmentation (Zhou et al. 2018; Liu et al. 2018).

Types of CNN:

CNNs are categorized based on the dimensionality of the input data and the sliding direction of the kernel:

Type Input Data Examples Kernel Movement
1D CNN Time-series, audio, ECG signals (Ige & Sibiya, 2024). Slides along one dimension (time/sequence).
2D CNN Grayscale/RGB images, medical X-rays (Taye, 2023). Slides along two dimensions (height and width).
3D CNN Videos, MRI/CT scans, 3D point clouds (Guida et al., 2021). Slides along three dimensions (height, width, and depth).

3D Convolutional Neural Networks:

3D CNNs extend the capabilities of 2D networks by adding a third dimension (depth or time) to the convolution operation (Ige & Sibiya, 2024). This allows the network to capture spatiotemporal or volumetric contexts that are often lost when processing 3D data as independent 2D slices (Baheti et al., 2023; Guida et al., 2021).This characteristic makes 3D CNN particularly well-suited for medical imaging modalities such as magnetic resonance imaging (MRI) and computed tomography (CT), where anatomical structures extend across multiple slices.

The 3D convolution is defined as:

\[ Y(i,j,k) = \sum_{m}\sum_{n}\sum_{p} X(i+m, j+n, k+p)\cdot W(m,n,p) + b \]

Where: X= Input 3D MRI volume W= 3D convolution kernel b= Bias Y= Output feature map i,j,k= Spatial voxel indices This operation extracts volumetric features across depth, height, and width.

Applications in Knee MRI Analysis

3D CNNs have demonstrated strong performance in musculoskeletal imaging, particularly in knee MRI analysis.

  • Tissue Segmentation:

    Liu et al. (2018) and Zhou et al. (2018) applied deep CNN-based frameworks for segmentation of knee joint anatomy, demonstrating improved accuracy in delineating cartilage and other tissues.

  • Meniscus and Cartilage Degeneration Detection:

    Pedoia et al. (2019) employed 3D CNN models to detect and stage degenerative morphological changes in meniscus and patellofemoral cartilage, highlighting the importance of volumetric feature extraction.

  • Osteoarthritis Classification:

    Guida et al. (2021) proposed a 3D CNN model for knee osteoarthritis classification using MRI, reporting enhanced performance compared to 2D approaches due to better preservation of spatial context. Similarly, Rani et al. (2024) demonstrated CNN-based classification for osteoarthritis severity assessment.

  • ACL Tear Detection: Bien et al. (2018) introduced MRNet, a deep learning framework for knee MRI diagnosis. More recent approaches incorporate advanced learning paradigms such as self-supervised learning for ACL tear detection (Aidarkhan et al., 2025).

  • Multimodal and Federated Learning Approaches: Emerging frameworks integrate 3D CNNs with federated and few-shot learning strategies to improve generalization across institutions (Goel, 2025). Oettl et al. (2025) further highlight the role of AI-driven multimodal systems in advancing orthopedic diagnostics.

Performance and Advantages:

3D CNNs provide significant benefits in medical imaging (Avesta et al., 2023; Guida et al., 2021):

  • Volumetric Context: They extract features from adjacent slices, detecting biomarkers (like cartilage degradation) that may be invisible in a single 2D image (Guida et al., 2021).

  • Higher Accuracy: Benchmark studies in brain and knee imaging show that 3D models generally achieve higher Dice scores and classification accuracy compared to 2D or 2.5D approaches (Avesta et al., 2023; Guida et al., 2021).

  • Efficiency in Convergence: 3D models can converge 20% to 40% faster during training than their 2D counterparts when dealing with volumetric data (Avesta et al., 2023).

Limitations and Assumptions

Despite their power, 3D CNNs face specific challenges (Avesta et al., 2023; Baheti et al., 2023):

  • Computational Cost: 3D models require significantly more computational memory (often up to 20 times more) compared to 2D models (Avesta et al., 2023).

  • Data Scarcity: Unlike 2D CNNs, which benefit from massive pre-trained datasets like ImageNet, 3D CNNs often suffer from a lack of large-scale pre-trained volumetric models, leading to potential stability issues with random weight initialization (Baheti et al., 2023).

Data Exploration and Visualization

Dataset Description – MRNet v1.0

For automated knee MRI diagnosis, we use the MRNet v1.0 dataset released by the Stanford Machine Learning Group**. The dataset contains knee magnetic resonance imaging (MRI) examinations collected at Stanford University Medical Center. It was introduced to support research in deep learning–based detection of common knee injuries.

The dataset consists of 1,370 knee MRI exams performed on different patients. Each exam includes multiple MRI slices captured in three anatomical planes: axial, coronal, and sagittal. For every exam, labels are provided for three diagnostic tasks:

Abnormality detection

ACL (Anterior Cruciate Ligament) tear detection

Meniscal tear detection

The labels were extracted from radiology reports and reviewed for research purposes. The dataset is widely used as a benchmark for medical image classification and multi-label learning tasks.

The dataset is typically split into: Training set: 1,130 exams

Validation set: 120 exams

Test set: 120 exams (held-out for evaluation in the original competition)

Data Definition

Unlike transactional datasets, MRNet is an exam-level medical imaging dataset. The data is organized as image volumes rather than tabular transaction records.

Attribute Type Description Study ID Nominal Unique identifier assigned to each MRI exam Plane Nominal MRI acquisition plane (Axial, Coronal, Sagittal) Image Slices Image Sequence Stack of 2D grayscale MRI slices forming a 3D volume Abnormal Binary (0/1) Indicates presence of any abnormality ACL Tear Binary (0/1) Indicates presence of ACL tear Meniscal Tear Binary (0/1) Indicates presence of meniscal tear

Each study contains:

Variable number of slices (typically 20–60 per plane)

Grayscale images

Stored in .npy format (NumPy arrays) ## Analysis and Results

Conclusion

  • Summarize your key findings.

  • Discuss the implications of your results.

References

Aidarkhan, Assanali, Azamat Mukhamediya, Almira Askhatova, Roza Beisembekova, Zhadyra Alimbayeva, Aiman Ozhikenova, Kassymbek Ozhikenov, and Prashant K Jamwal. 2025. “Self-Supervised Learning with BYOL for Anterior Cruciate Ligament Tear Detection from Knee MRI.” MethodsX, 103664. https://doi.org/10.1016/j.mex.2025.103664.
Astuto, Bruno, Io Flament, Nikan K. Namiri, Rutwik Shah, Upasana Bharadwaj, Thomas M. Link, Matthew D. Bucknor, Valentina Pedoia, and Sharmila Majumdar. 2021. “Automatic Deep Learning–Assisted Detection and Grading of Abnormalities in Knee MRI Studies.” Radiology: Artificial Intelligence 3 (3): e200165. https://doi.org/10.1148/ryai.2021200165.
Awan, Mazhar Javed, Mohd Shafry Mohd Rahim, Naomie Salim, Amjad Rehman, Haitham Nobanee, and Hassan Shabir. 2021. “Improved Deep Convolutional Neural Network to Classify Osteoarthritis from Anterior Cruciate Ligament Tear Using Magnetic Resonance Imaging.” Journal of Personalized Medicine 11 (11): 1163. https://doi.org/10.3390/jpm11111163.
Bien, Nicholas, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, et al. 2018. “Deep-Learning-Assisted Diagnosis for Knee Magnetic Resonance Imaging: Development and Retrospective Validation of MRNet.” PLoS Medicine 15 (11): e1002699. https://doi.org/10.1371/journal.pmed.1002699.
Goel, Chirag. 2025. “Federated Knee Injury Diagnosis Using Few Shot Learning.” Frontiers in Artificial Intelligence 8: 1589358. https://doi.org/10.3389/frai.2025.1589358.
Guida, Carmine, Ming Zhang, and Juan Shan. 2021. “Knee Osteoarthritis Classification Using 3D CNN and MRI.” Applied Sciences 11 (11): 5196. https://doi.org/10.3390/app11115196.
Liu, Fang, Zhaoye Zhou, Hyungseok Jang, Alexey Samsonov, Gengyan Zhao, and Richard Kijowski. 2018. “Deep Convolutional Neural Network and 3D Deformable Approach for Tissue Segmentation in Musculoskeletal Magnetic Resonance Imaging.” Magnetic Resonance in Medicine 79 (4): 2379–91. https://doi.org/https://doi.org/10.1002/mrm.26841.
Oettl, Felix C, Bálint Zsidai, Jacob F Oeding, Michael T Hirschmann, Robert Feldt, Thomas Tischer, Kristian Samuelsson, and ESSKA Artificial Intelligence Working Group. 2025. “Beyond Traditional Orthopaedic Data Analysis: AI, Multimodal Models and Continuous Monitoring.” Knee Surgery, Sports Traumatology, Arthroscopy 33 (6): 2269–75. https://doi.org/10.1002/ksa.12657.
Panwar, Punita, Sandeep Chaurasia, Jayesh Gangrade, and Ashwani Bilandi. 2025. “Early Diagnosis of Knee Osteoarthritis Severity Using Vision Transformer.” BMC Musculoskeletal Disorders 26 (1): 884. https://doi.org/10.1186/s12891-025-09137-2.
Pedoia, Valentina, Berk Norman, Sarah N Mehany, Matthew D Bucknor, Thomas M Link, and Sharmila Majumdar. 2019. “3D Convolutional Neural Networks for Detection and Severity Staging of Meniscus and PFJ Cartilage Morphological Degenerative Changes in Osteoarthritis and Anterior Cruciate Ligament Subjects.” Journal of Magnetic Resonance Imaging 49 (2): 400–410. https://doi.org/10.1002/jmri.26246.
Qiu, Xubin, Zhiwei Liu, Ming Zhuang, Dong Cheng, Chenlei Zhu, and Xiaoying Zhang. 2021. “Fusion of CNN1 and CNN2-Based Magnetic Resonance Image Diagnosis of Knee Meniscus Injury and a Comparative Analysis with Computed Tomography.” Computer Methods and Programs in Biomedicine 211: 106297. https://doi.org/10.1016/j.cmpb.2021.106297.
Rani, Suman, Minakshi Memoria, Ahmad Almogren, Salil Bharany, Kapil Joshi, Ayman Altameem, Ateeq Ur Rehman, and Habib Hamam. 2024. “Deep Learning to Combat Knee Osteoarthritis and Severity Assessment by Using CNN-Based Classification.” BMC Musculoskeletal Disorders 25 (1): 817.
Yeoh, Pauline Shan Qing, Khin Wee Lai, Siew Li Goh, Khairunnisa Hasikin, Yan Chai Hum, Yee Kai Tee, and Samiappan Dhanalakshmi. 2021. “Emergence of Deep Learning in Knee Osteoarthritis Diagnosis.” Computational Intelligence and Neuroscience 2021 (1): 4931437. https://doi.org/10.1155/2021/4931437.
Zhou, Zhaoye, Gengyan Zhao, Richard Kijowski, and Fang Liu. 2018. “Deep Convolutional Neural Network for Segmentation of Knee Joint Anatomy.” Magnetic Resonance in Medicine 80 (6): 2759–70.