Towards Effective and Interpretable Deep Learning Applications for Microscopic Medical Imaging

点击翻译按钮获取中文摘要

Deep learning has quickly evolved to solve difficult problems in many domains such as computer vision, particularly for problems where large amounts of data are available. For many medical applications, though, the data available tend to be small to medium size data sets. Further, these data sets typically are incomplete, contain entry errors, and unbalanced in terms of target classifications. For medical imaging data, beyond the basic complex nature of anatomical structures, there are challenges of noise from the collection process and the presence of imaging artifacts. The long-term objective of our research is to develop robust, accessible, rapid, transparent and accurate computer-aided medical diagnosis and grading tools to provide an automatic independent second opinion, to help avoid errors/oversights, to clinicians. The proposed research in our proposal aims to introduce or adapt new deep learning approaches to solve problems related to classification, segmentation, and interpretability for small- or medium-scale medical data. The classification problems considered are primarily related to the development of Computer Aided Diagnosis (CAD) systems, particularly for cytological images of fine needle biopsies for breast cancer, where we seek to improve accuracy and robustness to achieve clinically acceptable performance. We will continue our research on leveraging the progression of a disease over time to improve diagnosis. To create augmented data for deep learning training, we will study the creation of images and sequences of medical images by adapting our recent work on generative adversarial networks. For noisy medical data, we will seek to improve training in the presence of such data using local intrinsic dimensionality to locally evaluate the trained model. Our recent work on Continual Learning will be extended to improve the reliability of CAD systems to allow online learning of shifts in data distributions. Finally, we will our continue work on Meta-Learning by extending recent work on unsupervised Few-Shot Meta-Learning to radiomic applications. We will continue our work on improved nuclear segmentation algorithms for optical microscopic images. In addition to trying new deeper architectures, we will also explore the design of new loss functions for segmenting medical images, exploring the development of techniques beyond optimization approaches such as deep supervision, and the development of algorithmic learning rate schedulers. For deep learning models for malignancy diagnosis or grading, one major difficulty in these systems being approved for clinical use is the lack of interpretability of the decisions made by these models. To interpret the decisions made by deep learning models, we are studying correlations between handcrafted features and deep learning model features, the image retrieval strategies for finding similar images from past patients, and the generation of counterfactual visual explanations.