- SSR 2019 Podium Presentation
- Researchers funded by RSNA, Institutional Sarcoma Grant, and NIH Health Grants
- All myxoid tumors can be differentiated from each other using fluorescent in-situ hybridization (FISH) or immunohistochemical markers, except for myxomas and myxofibrosarcomas. Myxomas are benign and histologically bland, whereas myxofibrosarcomas are malignant and histologically heterogenous. Because of the histological heterogeneity, low grade myxofibrosarcomas may be mistaken for myxomas on core needle biopsies.
- In this work we evaluated the performance of T1-weighted signal intensity (T1SI), tumor volume, and radiomic features extracted from magnetic resonance imaging (MRI) to differentiate myxomas from myxofibrosarcomas.
- Our results demonstrated that radiomic features extracted from T1-weighted sequences can provide better discriminative information in distinguishing myxoma from myxofibrosarcomas when compared to T1-weighted signal intensity values and tumor volume.
- My favorite part of carrying out this research was our regular meetings with our data scientist colleagues to discuss the project. The meetings were very engaging, and I learned great deal about AI/ML in the process. These skills proved very useful in subsequent research projects!
Radiomics-based Machine Learning Distinguish
Lipomas From Atypical Lipomatous
- SSR 2022 Podium Presentation
- Manuscript in final stages of preparation
- Lipomatous tumors are commonly encountered in clinical practice. While it is generally not challenging to distinguish high-grade liposarcomas from benign lipomas at imaging, it is more difficult to differentiate benign lipomas from Atypical Lipomatous Tumors (ALT)/Well-differentiated Liposarcomas (WDLPS). In the literature, several MRI discriminators have been described to differentiate these tumors, however, considerable overlap remains which makes accurate distinction more challenging for radiologists. This differentiation is clinically important to ensure that patients receive the correct treatment and follow-up, when applicable.
- In this work we developed a classifier model using an MRI-based radiomics approach to differentiate lipomas from ALT/WDLPS on MRI examinations.
- Our research demonstrated that radiomics is a promising, non-invasive approach for differentiating lipomas from ALT/WDLPS.
- This project was conducted during my fellowship training year. Transitioning to a new institution for fellowship training posed unique challenges such as finding new machine learning expert researchers with similar interests and learning to navigate and query data from a new radiology information system. After a few inquiries, I was connected to the Machine Learning for Medical Imaging cohort at UW-Madison in which I found outstanding collaborators to help carry out our research. Collaboration is key to successful research!
Radiomics Analysis Workflow
My Research - Final Thoughts
- I would like to thank the Society of Skeletal Radiology for the opportunity to present my research at the annual meeting. Podium presentations and the subsequent discussions at the annual meeting have contributed to new ideas for future projects, conceptualizing our work in new and often improved ways, and they have also served to address relevant research questions prior to manuscript publication.
- Image-based radiomic analysis of tumors has considerable potential to improve decision-making processes in daily clinical practice. Radiomics utilizes advanced algorithms to extract quantitative imaging features from medical images that are imperceptible to our eyes when interpreting cases at the workstation. When coupled with machine learning analysis, it can detect even higher dimension patterns. Our research in this domain has demonstrated that radiomics can provide non-invasive discriminative information in distinguishing between myxomas and myxofibrosarcomas, and between lipomas and well-differentiated liposarcomas. However, future studies are needed using larger, external populations for model validation prior to integration into our daily clinical workflows.