Abstract #2657

Generalizability of Deep-Learning Segmentation Algorithms on Independent Datasets for Measuring T2 Relaxation Times

Andrew Schmidt¹, Arjun D Desai¹, Lauren E Watkins², Hollis Crowder³, Elka Rubin¹, Valentina Mazzoli¹, Quin Lu⁴, Marianne S Black^1,3, Feliks Kogan¹, Garry E Gold^1,2, Brian A Hargreaves^1,5, and Akshay S Chaudhari¹

¹Radiology, Stanford University, Palo Alto, CA, United States, ²Bioengineering, Stanford University, Palo Alto, CA, United States, ³Mechanical Engineering, Stanford University, Palo Alto, CA, United States, ⁴Philips Healthcare North America, Gainesville, FL, United States, ⁵Electrical Engineering, Stanford University, Palo Alto, CA, United States

Automated segmentation using deep learning can potentially expedite segmentation tasks. However, the generalizability of such algorithms on new unseen datasets is unknown. To test this generalizability, we used a knee segmentation algorithm trained on Osteoarthritis Initiative double-echo steady-state (DESS) datasets to segment cartilage from quantitative DESS datasets from three independent studies. We compared manual-automatic segmentation accuracy and the resultant qDESS T₂ map variations. These results quantitatively demonstrate that a deep learning network trained on a single dataset does not generalize with a high accuracy to additional datasets even with similar image characteristics, and that additional fine-tuning may be needed.

This abstract and the presentation materials are available to members only; a login is required.

Join Here