Project Details
Description
Clinically Magnetic Resonance Imaging (MRI) is widely utilized to assess the internal disorders of the knee, especially in tissues such as cartilage, bone, menisci and ligaments. Since the joint shape and tissue degenerations have a strong influence on knee related injuries and diseases, therefore non-invasive characterization of these features by MRI is valuable for disease diagnosis as well as planning therapeutic procedures. These investigations from MRI using manual delineation of knee tissue is time consuming and can be impractical when processing large cohorts in routine clinical applications. There is a pressing need to develop automatic machine intelligence-based imaging tools offering predictive outcomes that provide morphological and compositional insights to facilitate the understanding of disease as well as interpret insights from images timely and effectively. To this end, we propose to develop artificial intelligence (AI) based tool that can process a large cohort of MRI scans and generate an assessment record of knee anatomy with various morphological features along with 3D visualization of knee bone and tissue’s structure. The assessment record can further be processed to facilitate the practitioners to make decisions along with initial indications of predictive elements.
Deep convolutional neural networks (CNNs) have gained much popularity in computer vision society for cracking complex problems, however, in knee MRI the availability of suitable labelled data is a key factor that limits its direct application. Another common issue with supervised methods is their adaptability to newly acquired data in different domains. Usually, the data set used to train CNNs algorithms to come from the public datasets that are acquired in specific control settings, when it comes to testing the model on unseen unique samples from other domains (with different vendors and different acquisition protocols), the CNNs methods fail to adopt those changes and results in incomplete segmentation. The major concern in developing AI-based system in knee MRI is to mitigate the generalization problem or domain adaptation that arises with qualitative diversity among examinations and vary in imaging protocols, subjects, and hardware.
We argue that the self-supervised method with domain adaptation can be useful to solve this issue and benefit the knee MRI segmentation with better generalization abilities. To this end, we present an adversarial learning-based self-supervised segmentation approach that addresses the above-mentioned challenges in an unsupervised manner and provides a reliable estimate of favorable knee segmentation without user intervention. In our preliminary study, we have implemented and tested our idea on knee MRI data that features the knee MRI scans from the separate domains with different imaging protocols as well as vendors. The initial segmentation results show that the proposed method addresses the problem of tissue segmentation quite well. However, further experimental study is required to implement full pipeline and test on benchmark datasets as well as independent datasets to validate the segmentation. The segmentation from knee tissues serves the initial purpose of tissue region segmentation with domain adaptation perceptive. For the assessment module, we aim to establish model assisted interpretation. The module will be capable of processing the segmentation and producing the relevant clinical information such as (thickness measurements in 2D and 3D, abnormalities in meniscus such as tears, degradation in cartilages such as cartilage loss, etc). The information collected will be used to establish the domain invariant assessment along with 3D rendering of the knee structures. The automated assessment will be further compared with the benchmark gold standards, and thorough validation procedures will be conducted. Improvements will be made by exploring further key clinically relevant features by extracting and manslaying the radiomics features. The outcomes will be presented in terms of a compact tool that possess both qualitative as well as quantitative assessments of the knee MRI independent of the acquired domain.
Deep convolutional neural networks (CNNs) have gained much popularity in computer vision society for cracking complex problems, however, in knee MRI the availability of suitable labelled data is a key factor that limits its direct application. Another common issue with supervised methods is their adaptability to newly acquired data in different domains. Usually, the data set used to train CNNs algorithms to come from the public datasets that are acquired in specific control settings, when it comes to testing the model on unseen unique samples from other domains (with different vendors and different acquisition protocols), the CNNs methods fail to adopt those changes and results in incomplete segmentation. The major concern in developing AI-based system in knee MRI is to mitigate the generalization problem or domain adaptation that arises with qualitative diversity among examinations and vary in imaging protocols, subjects, and hardware.
We argue that the self-supervised method with domain adaptation can be useful to solve this issue and benefit the knee MRI segmentation with better generalization abilities. To this end, we present an adversarial learning-based self-supervised segmentation approach that addresses the above-mentioned challenges in an unsupervised manner and provides a reliable estimate of favorable knee segmentation without user intervention. In our preliminary study, we have implemented and tested our idea on knee MRI data that features the knee MRI scans from the separate domains with different imaging protocols as well as vendors. The initial segmentation results show that the proposed method addresses the problem of tissue segmentation quite well. However, further experimental study is required to implement full pipeline and test on benchmark datasets as well as independent datasets to validate the segmentation. The segmentation from knee tissues serves the initial purpose of tissue region segmentation with domain adaptation perceptive. For the assessment module, we aim to establish model assisted interpretation. The module will be capable of processing the segmentation and producing the relevant clinical information such as (thickness measurements in 2D and 3D, abnormalities in meniscus such as tears, degradation in cartilages such as cartilage loss, etc). The information collected will be used to establish the domain invariant assessment along with 3D rendering of the knee structures. The automated assessment will be further compared with the benchmark gold standards, and thorough validation procedures will be conducted. Improvements will be made by exploring further key clinically relevant features by extracting and manslaying the radiomics features. The outcomes will be presented in terms of a compact tool that possess both qualitative as well as quantitative assessments of the knee MRI independent of the acquired domain.
| Status | Finished |
|---|---|
| Effective start/end date | 1/01/23 → 30/06/24 |
Keywords
- Knee MRI Segmentation, Deep Learning, Self supervised learning, VISTA
Fingerprint
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.