Dideolu Dawodu (Author and Presenter) Acknowledgments: Dr. Garry Gold (Primary Mentor), Andrew Schmidt (Primary Mentor), Arjun Desai (DOSMA Technical Support), Elka Rubin (Data Acquisition and Logistics)

Purpose: Quantitative MRI can be a valuable, noninvasive tool for early tracking of physiological changes in the joint health of at-risk populations, such as athletes. A crucial preliminary step, segmentation, involves accurately demarcating the anatomical area of interest on the visualization scan (MRI, in this case); a human annotator generally does it, but it is incredibly time-consuming. For this reason, this project explores if machine models can significantly reduce the preparation time of segmentation, as previous research has shown, without considerably sacrificing accuracy.

Methods: The data was obtained from the researcher mentioned above, Elka Rubin. Within this dataset, bilateral knee scans (again, MRI) were taken for both swimmers (control group) and basketball players (experimental group) and then manually segmented. The statistical and visual accuracy of the machine models known as the quantitative double echo steady state data (qDESS) and the International Workshop Osteoarthritis Initiative (IWOAI) were tested against these manually segmented scans using the computer software known as the Deep Open-Source MRI Analysis (DOSMA) and the coding program known as Python.

Results: The qDESS model did notably better than the IWOAI model in terms of the Dice Similarity Coefficient (DSC) scores, which directly measure segmentation accuracy. Additionally, higher Concordance Correlation Coefficient (CCC) scores for the qDESS model indicate increased agreement with the manually segmented images.

Interpretations: It is standard practice to create and train machine models to the dataset they will be segmenting. The qDESS and IWOAI models were developed and pre-trained on broad-based datasets unrelated to these athlete populations. So, while DSC and CCC scores were not indicative of high statistical accuracy for either machine model, further investigations can explore how to make machine models more adaptable.

Conclusions: Automated segmentation could become the golden standard if machine models can be fine-tuned to perform well in segmentation accuracy and speed without prior priming to any particular dataset population.