(Thursday, 7/16/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room: Track 5
Purpose: Supervised deep-learning methods require a large amount of labeled data for model training. Data-labeling in medical imaging practice is tedious, expensive, and prone to subjective errors. Transfer learning (TL) and domain adaption (DA) methods have been applied to address this issue. However, the severity of source-target domain shifts and the number of the available labeled data affect their performance. This is the first quantitative study to provide general guidance on using TL and/or DA methods for medical imaging applications.
Methods: Learning a CNN-based numerical observer with computer-simulated images for a binary signal detection task was employed as an example to conduct the quantitative analysis. DA models were trained to adapt a well-trained source observer (SO) to SODA observer for a target domain task. TL with varied number of labeled target domain samples was employed to SO and SODA to build TL- and DA-TL-based observers for target domain tasks respectively. The performance of SO, SODA, TL, and DA-TL-based observers were analyzed and compared.
Results: The AUC increases for TL and DA-TL along with increased number of labeled target data. DA improves the performance of SO employed to target domains without using labeled target-domain data. AUC decreases for all methods when domain shift increases, SO and DA-TL decreases the fastest and slowest, respectively. When domain shift increases, DA-TL yields more obviously better performance than other methods. Both TL and DA-TL can improve the performance of SO adopted to target domains containing limited labeled target data. This is more obvious on target domains with larger shifts.
Conclusion: Appropriate employing DL and DA methods for the tasks in target domains should consider the severity of domain shifts and the number of available labeled target domain data. The quantification strategy employed in this study can be extended to other DL and DA methods.
Funding Support, Disclosures, and Conflict of Interest: This research was supported in part by NIH awards EB020604, EB023045, NS102213, EB028652, R01CA233873, R21CA223799, and NSF award DMS1614305.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
