Domain-specific deep learning with environmental and medical applications

Abstract

Deep learning has achieved unprecedented success in various fields ranging from image recognition to natural language processing. However, despite their prowess, deep learning models can encounter issues related to data inefficiency, overfitting, lack of interpretability and sensitivity to dataset shifts. Furthermore, these models often function as "black boxes", providing little insight into the internal workings of the model or the decision-making process. To address these challenges, Domain-specific Deep Learning has risen as a growing trend. Domain-specific Deep Learning refers to the integration of domain-specific knowledge in deep learning models. Problem-specific insights can be utilized to either enhance the deep learning model or facilitate the training process. This approach aims to leverage the expressivity and learning capacity of neural networks while benefitting from the precision, robustness and interpretability provided by domain-specific knowledge. This study proposes several different approaches for incorporating domain-specific knowledge into deep learning models. We provide examples of how these methods can be applied in the fields of biomedical and environmental science, illustrating their wide-ranging benefits. Chapter 3 integrates domain knowledge with a deep learning model by acting on the feature space. Our proposed Deep-AIR model fills the existing research gap in air pollution modelling by exploiting domain-specific features (including Air Pollution, Weather, Urban Morphology, Transport and Time-sensitive features), and by employing a hybrid CNN-LSTM structure to capture the spatial-temporal features and 1×1 convolution layers to enhance the learning of temporal and spatial interaction. Chapter 4 injects domain knowledge into a deep learning model by exploiting regularisation and post-processing schemes with prior insights. We develop a deep learning-based 3D cell segmentation pipeline, 3DCellSeg, to address challenges in cell segmentation. A custom loss function (3DCellSeg Loss) is introduced to tackle the clumped cell problem, and an efficient touching area-based clustering algorithm (TASCAN) is developed to separate 3D cells from foreground masks. Chapter 5 explores a deeper combination of knowledge and network, by guiding the construction of a neural network with domain-specific knowledge. We develop a novel hierarchical graph attention network to identify highly correlational GMs (HGMs) of PTSD. Domain-specific knowledge, including somatic mutations, genes, PTSD pathways and their correlations have been incorporated into the graph structures. Our study is carefully guided by prominent PTSD literature and clinical experts within the field; highly salient HGMs generated from our model are further verified by existing PTSD-related authoritative medical journals. Our study illustrates the utility and significance of a hybrid approach, integrating both AI and expert-guided/domain-specific knowledge for the thorough identification of biomarkers. Our work is among the first to systematically explore different approaches to integrate domain-specific knowledge with deep learning models, emphasising not only the theoretical contributions but also the practical implications. Our experimental results demonstrate the improvement of deep learning models on both inference precision, and interpretability and robustness. This study underpins the imperative of converging the power of deep learning with the richness of domain-specific knowledge, thus facilitating the advancement of more sophisticated, effective and trustworthy AI systems.