Date of Award
Open Access Thesis
Medical Doctor (MD)
Background: As the quantity and complexity of oncological data continue to increase, machine learning (ML) has become an important tool in helping clinicians better understand malignancies and provide personalized care. Diagnostic image analysis, in particular, has benefited from the advent of ML methods to improve image classification and generate prognostic information from imaging collected in routine clinical practice [1-3]. Deep learning, a subset of ML, has especially achieved remarkable performance in medical imaging, including segmentation [4, 5], object detection, classification , and diagnosis .
Despite the notable success of deep learning computer vision models on oncologic imaging data, recent studies have identified notable weaknesses in deep learning models used on diagnostic images. Specifically, deep learning models have difficulty generalizing to data that was not well represented during training. One potential solution is the use of domain adaptation (DA) techniques, which improve the generalizability of a deep learning model trained on one domain to better generalize to data of a target domain.
Techniques: In this study, we explain the efficacy of four common DA techniques – MMD, CORAL, iDANN, and AdaBN - used on deep learning models trained on common diagnostic imaging modalities in oncology. We used two datasets of mammographic imaging and CT scans to test the prediction accuracy of models using each of these DA techniques and compared them to the performance of transfer learning.
Results: In the mammographic imaging data, MMD, CORAL, and iDANN increased the target test accuracy for all four domains. MMD increased target accuracies by 3.6 - 45%, CORAL by 4- 48%, and iDANN by 1.5-49.4%. For the CT scan dataset, MMD, CORAL, and iDANN increased the target test accuracy for all domains. MMD increased the target accuracy by 2.0 – 13.9%, CORAL by 2.4 - 15.8%, and iDANN by 2.0 – 11.1%. in both the mammographic imaging data and the CT scans, AdaBN performed worse or comparably to transfer learning.
Conclusion: We found that DA techniques significantly improve model performance and generalizability. These findings suggest that there’s potential to further study how multiple DA techniques can work together and that these can potentially help us create more robust, generalizable models.
Veira, Juliana, "Improving Cancer Classification With Domain Adaptation Techniques" (2023). Yale Medicine Thesis Digital Library. 4203.
This Article is Open Access