Strategy for Improving Generalizability of Medical Imaging Artificial Intelligence Model
- 주제(키워드) Deep Learning , Medical Imaging , Style Transfer
- 주제(DDC) 570
- 발행기관 아주대학교
- 지도교수 박래웅
- 발행년도 2022
- 학위수여년월 2022. 2
- 학위명 박사
- 학과 및 전공 일반대학원 의생명과학과
- 실제URI http://www.dcollection.net/handler/ajou/000000031797
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Recent developments in the field of artificial intelligence have helped develop high-performance artificial intelligence models in the medical imaging field. However, when applying these models in the actual medical site, performance declines are frequently observed. This is because the heterogeneity of medical imaging data is greater than that of the data obtained from other fields; this heterogeneity is a major obstacle to the growth of the artificial intelligence software market for medical imaging. Until now, to cope with domain shifts associated with the use of artificial intelligence in medical imaging, scientists have applied the color normalization and data-augmentation techniques to the training data, or the developed model has been finely adjusted to the test data. However, these techniques had limitations, such as the artificial adjustment of learning data diversity and the lack of data labels to fine-tune the developed model. In this paper, we use a generative model-based style-transfer technique to prevent any performance decline in the actual application stage of the artificial intelligence model for medical imaging analysis. When applying the artificial intelligence model to the actual medical sites, we use the generative model along with the artificial intelligence model developed by learning the medical image. The generative model adopted in this paper is CycleGAN, which converts landscape photographs into paintings of the artist Gogh and conversely. Using this model, the style of the actual medical field images that the artificial intelligence model should predict is changed to the style of the images learned by the artificial intelligence model. The analysis revealed that the artificial intelligence model showed significantly lower performance when predicting images of the actual field with styles different from the image that they learned. However, in images that passed through the generative model and were changed to a style familiar to the artificial intelligence model, the performance was restored. We confirmed that performance improvement of the proposed model was greater than that of the color-normalization methods used in the pathology field. Besides pathology images, brain magnetic resonance imaging (MRI) data in the field of radiology confirmed that changing the style improved the generalizability of the model. In addition, the mechanism of performance improvement was inferred by analyzing the image characteristics before and after the application of CycleGAN and the changes in the feature values extracted from the images. The results of this paper prove that the use of the style-transfer technique based on the generative model can be an effective strategy to prevent performance degradation because of data heterogeneity when the medical artificial intelligence model is applied.
more목차
I. Introduction 1
A. Motivations 1
B. Generative Adversarial Network for Style Transfer between Domains 6
C. Dissertation outline 10
II. First Study in Pathology Domain 12
A. Introduction 12
B. Materials and Methods 16
1. Dataset 18
2. Classifier development 21
3. Leveraging the style-transfer method 22
4. Applying the style-transfer in different type of cancer 23
5. Information of software and hardware 24
6. Ethics statement 24
C. Results 25
1. Internal validation 25
2. Classifier’s performance 27
3. Effects of style transfer on performance 38
D. Discussion 40
III. Second Study in Radiology Domain 42
A. Introduction 42
B. Materials and Methods 44
1. Patient population 44
2. Pathological diagnosis 46
3. MRI protocol 46
4. Image preprocessing and radiomic feature extraction 48
5. Radiomics model construction 51
6. CycleGAN application 53
7. Evaluation of the effect of CycleGAN: Fréchet Inception Distance and t-Distributed Stochastic Neighbor Embedding 56
C. Results 57
1. Performance of the classifier for the original external validation and CycleGAN style-transferred external validation images 59
2. Evaluation of the effect of CycleGAN: FID and t-SNE 69
D. Discussion 74
IV. Conclusion 78
References 80
국문요약 96
