# 세줄요약 #
- 최근 뇌영상 처리(Neuroimaging Techniques) 기법들의 빠른 발전과 MRI, PET 등 다양한 종류에서 방대한 양의 뇌영상 데이터들이 나오면서, 딥러닝을 이용하여 알츠하이머 질병의 조기 발견과 자동 분류를 해주는 진단 모델들이 연구 및 개발되고 있다.
- 이 논문에서는 그러한 알츠하이머 진단 모델들의 기존 연구들을 평가하고 정리하기 위해 총 16개의 논문들을 리뷰하였으며, 그 중 4개는 딥러닝과 함께 전통적인 머신러닝 기법을 사용한 연구들이고, 나머지 12개는 오직 딥러닝 기법만 사용한 연구들이다.
- 딥러닝 기법은 기존의 전통적인 머신러닝 기법에서 전제되는 특성 추출을 위한 복잡하고 어려운 뇌 영상 처리 기법들을 적용할 필요가 없기에, 현재 다양한 종류의 뇌영상(MRI, PET 등등)들에서 알츠하이머 진단 모델 개발에 사용되고 있으며 그 성능 또한 점점 향상되고 있다.
# 상세리뷰 #
- Introduction
- AD is an irreversible, progressive brain disorder marked by a decline in cognitive functioning with no validated disease modifying treatment (De strooper and Karran, 2016).
- Thus, a great deal of effort has been made to develop strategies for early detection, especially at pre-symptomatic stages in order to slow or prevent disease progression (Galvin, 2017; Schelke et al., 2018).
- Rapid progress in neuroimaging techniques has made it challenging to integrate large-scale, high dimensional multimodal neuroimaging data.
- Therefore, interest has grown rapidly in computer-aided machine learning approaches for integrative analysis.
- Hybrid method: well defined features influence performance results
- High accuracy result, but concern about overfitting!
- Suk and Shen, 2013: SAE + SVM, first paper used machine learning method, 95.9% acc.
- Suk et al., 2015: Best performance paper in this review, AD/CN classification accuracy = 98.8%
- Deep learning identifies optimal features automatically form data
- Diagnostic classification without human intervention
- Deep learning aimed at faster analysis with better accuracy than human practitioners.
- Liu et al., 2014: first paper used only deep learning method(SAEs + softmax regression layer), 87.8% acc.
- Choi & Jin, 2018: first used 3D CNN to multimodal PET images, 96% acc.
- Deep learning have several issues need to be addressed, Transparency & Reproducibility
- Transparency: deep learning algorithm have uncertainties and complexity, so the mathematical background is difficult to explain.
- Reproducibility: The uncertainty of the configuration and the randomness involved in the training procedure may make it difficult to reproduce the study and achieve the same results.
- Future direction: The expansion of 2D CNN into 3D CNN, Used multimodal neuroimages, GAN applied to generating synthetic medical images for data augmentation, Reinforcement learning
* Reference: Jo T, Nho K and Saykin AJ (2019) Deep Learning in Alzheimer’s Disease: Diagnostic Classification and Prognostic Prediction Using Neuroimaging Data. Front. Aging Neurosci. 11:220. doi: 10.3389/fnagi.2019.00220
728x90
728x90
댓글