MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Generalizability Issue of Deep Learning Models in Medicine and Its Potential Solutions: Illustrated with CBCT to CT Image Conversion

X Liang*, D Nguyen, S Jiang, UT Southwestern Medical Center, Dallas, TX

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose:
Generalization is a concern when applying a deep learning (DL) model trained on one dataset to other datasets. It is very challenging to demonstrate a DL model’s generalizability in an efficient and sufficient way before it can be implemented in clinical practice. In this work we will demonstrate the generalizability problem first, and then explore potential solutions based on transfer learning by using cone beam computed tomography (CBCT) to CT image conversion task as the testbed.

Methods:
We have collected 6 CBCT datasets for experiments. The datasets include: head and neck (H&N) and prostate from Varian TrueBeam/VitalBeam, prostate and pancreas from Elekta Versa, prostate and cervix from Elekta Agility. We divided the datasets into source domain and target domain, where source domain includes H&N data and target domain includes all the others. CycleGAN is used to train 4 models: source model, target model, combined model, and adapted model. Source model is only trained on the source domain and directly applied to a target domain without any modification. Target model is trained on a target domain from scratch. Combined model is trained on both source and target domain from scratch. Adapted model fine-tunes the trained source model to a target domain. Source model demonstrates the generalizability problem and the other models is used to investigate potential solutions.

Results:
We found that source model does perform badly when applied to a dataset coming from different machines, showing the significant generalization problem. Among three potential solutions, adaptive model has the best performance among all target domain datasets.

Conclusion:
Generalizability could be a significant problem when applying a trained DL model to different datasets. Adapting the trained source model using the fine-tuning strategy to an unseen target domain dataset is a viable and easy way to clinically implement the DL model.

Funding Support, Disclosures, and Conflict of Interest: This research is funded by Varian Research Grand.

Keywords

Cone-beam CT, Image Correlation

Taxonomy

IM/TH- Cone Beam CT: Scatter correction/ reduction

Contact Email