AI Segmentation in Clinically Relevant Complex Orthopedic Cases - A Targeted Study on Developmental Dysplasia of the Hip

Description

The presentation from the ISTA 2024 conference held in Nashville, gives a detailed discussion on advancements in AI-driven image segmentation techniques used in medical imaging, particularly focusing on challenging cases in orthopedic contexts. The speaker emphasizes the transition from manual segmentation, which is time-consuming, to AI automation that significantly reduces processing time to mere minutes. The discussion highlights issues with AI output, especially when encountering edge cases where results can deviate from expectations.



The talk specifically examines a dataset involving nine patients with dysplastic hips, comparing the automated AI segmentation results to manual segmentation done by skilled operators. Quantitative metrics such as Dice coefficient, Hausdorff distance, and mean absolute difference are presented to assess the performance of AI segmentation against manual methods. The results indicate high accuracy in common cases; however, significant errors occurred in complex regions, such as the acetabular rim and femoral heads, especially affected by osteophytes.



Lastly, the speaker details ongoing refinements to the AI model aimed at improving accuracy, particularly in challenging areas, indicating that the latest model achieves an average accuracy within half a millimeter. The model is made available for external evaluation, aiming to gather more data for continuous improvement.

Specialties

Conferences