Working principles for training neural networks with highly incomplete dataset: vanilla (upper panel) vs GapNet (lower panel) (Image by Yu-Wei Chang.)GapNet: Neural network training with highly incomplete datasets
Presentation in group meeting of Prof. Michael Strano, Department of Chemical Engineering, Massachusetts Institute of Technology, USA and DiSTAP, Singapore-MIT Alliance for Research and Technology, Singapore. Date: 23 February 2024
Neural network training requires complete data. We have introduced GapNet, which can train neural networks with incomplete data, using medical data. This approach can be generalized for integrating spectrum data across different frequency ranges, allowing the neural network to combine important information from diverse spectrum datasets.
Logo of the Gun and Bertil Stohne’s Foundation. (Image from the Foundation’s website.)
Yu-Wei Chang received one of the Gun and Bertil Stohnes Foundation Prizes for PhD students, with his recent research focusing on deep learning analysis of longitudinal tau pathology. The price consists in 100000 SEK given to one – or shared between two – student(s) at a Swedish university.
The Gun and Bertil Stohnes Foundation awards this prize to research projects in geriatrics that the Board deems of exceptional interest and value.
Anna Canal Garcia, from Karolinska Institutet and supervised by Prof. Joana B. Pereira, is the other recipient of this award. Anna’s research focuses on the intricate multilayer network analysis of brain neuroimaging data.
Opponent Saikat Chatterjee (on Zoom), Yu-Wei Chang (left), and PhD co-supervisor Joana B. Pereira (right). (Photo by P.-J. Chien.)Yu-Wei Chang completed the first half of his doctoral studies and he defended his half-time on the 3rd of November 2023.
The presentation was conducted in a hybrid format, with part of the audience present in the Nexus room and the remainder connected through Zoom. The seminar comprised a presentation covering both his completed and planned projects, followed by a discussion and questions posed by his opponent, Prof. Saikat Chatterjee.
The presentation commenced with an overview of his concluded projects. The first project involves handling incomplete medical datasets using neural networks and is published in ‘Machine Learning: Science and Technology.‘ It then transitioned to his second project, focusing on the development of software for brain connectivity analysis using multilayer graphs and deep learning. The corresponding repository is accessible on GitHub. In the final segment, he outlined the proposed continuation of his PhD, discussing an ongoing project centered around the deep learning analysis of longitudinal brain neural imaging data.
The proposed method enables accurate synthesis of longitudinal tau pathology. (Image by Y.-W. Chang.)Synthesizing tau pathology from structural brain imaging using deep learning Yu-Wei Chang, Giovanni Volpe, Joana B Pereira Date: 22 August 2023 Time: 10:15 AM PDT
In vivo tau-positron emission tomography (PET) is crucial for determining the stage of Alzheimer’s disease (AD). However, this method is expensive, not widely available, and exposes patients to ionizing radiation, which poses a carcinogenic risk. To address this issue, I’ll present our proposed method, a deep-learning synthesis approach for follow-up tau-PET brain images from baseline tau-PET images using a generative adversarial network (GAN). This technique has the potential to provide valuable insights into the progression of AD, the effectiveness of new treatments, and more accurate diagnosis of the disease.
Working principles for training neural networks with highly incomplete dataset: vanilla (upper panel) vs GapNet (lower panel) (Image by Y.-W. Chang.)
Training of neural network with incomplete medical datasets
Yu-Wei Chang
Neural network training and validation rely on the availability of large high-quality datasets. However, in many cases, only incomplete datasets are available, particularly in health care applications, where each patient typically undergoes different clinical procedures or can drop out of a study. Here, we introduce GapNet, an alternative deep-learning training approach that can use highly incomplete datasets without overfitting or introducing artefacts. Using two highly incomplete real-world medical datasets, we show that GapNet improves the identification of patients with underlying Alzheimer’s disease pathology and of patients at risk of hospitalization due to Covid-19. Compared to commonly used imputation methods, this improvement suggests that GapNet can become a general tool to handle incomplete medical datasets.
Deep-learning-detected tau deposition (color in orange) for Alzheimer’s Disease. (Image by Y.-W. Chang.)Deep-learning analysis in tau PET for Alzheimer’s continuum
Yu-Wei Chang, Giovanni Volpe, Joana B Pereira Submitted to SPIE-ETAI Date: 24 August 2022 Time: 16:40 (PDT)
Previous studies have suggested that Alzheimer’s disease (AD) is typically characterized by abnormal accumulation of tau proteins in neurofibrillary tangles. This is usually assessed by measuring tau levels in regions of interest (ROIs) defined based on previous post-mortem studies. However, it remains unclear where this approach is suitable for assessing tau accumulation in vivo across the different stages of individuals. This study employed a data-driven deep learning approach to detect tau deposition across different AD stages at the voxel level. Moreover, the classification performance of this approach on distinguishing different AD stages was compared with the one using conventional ROIs.
Working principles for training neural networks with highly incomplete dataset: vanilla (upper panel) vs GapNet (lower panel) (Image by Y.-W. Chang.)Neural network training with highly incomplete medical datasets
Yu-Wei Chang, Laura Natali, Oveis Jamialahmadi, Stefano Romeo, Joana B Pereira, Giovanni Volpe Submitted to SPIE-ETAI Date: 24 August 2022 Time: 08:00 (PDT)
Neural network training and validation rely on the availability of large high-quality datasets. However, in many cases, only incomplete datasets are available, particularly in health care applications, where each patient typically undergoes different clinical procedures or can drop out of a study. Here, we introduce GapNet, an alternative deep-learning training approach that can use highly incomplete datasets without overfitting or introducing artefacts. Using two highly incomplete real-world medical datasets, we show that GapNet improves the identification of patients with underlying Alzheimer’s disease pathology and of patients at risk of hospitalization due to Covid-19. Compared to commonly used imputation methods, this improvement suggests that GapNet can become a general tool to handle incomplete medical datasets.