End-to-End eXplainable Model Designs in Deep Learning for Computational Pathology

点击翻译按钮获取中文摘要

Deep learning models in computer vision are highly specialized for natural imaging domain. The knowledge design has been favored for such image type selection, but little is done for new arrivals such as computational pathology (i.e. a variant of medical imaging domain). Most models for image representation in computational pathology are the bi-products from computer vision domain-at most minor adjustments are made for application. Such blind transfer of model architectures and optimization algorithms are not necessarily the best practice for computational pathology. There are key differences between natural images and pathology images and because of that, this leads to issues such as the lack of numerical efficiency for handling bigdata in pathology, eXplainability, and suitability of AI models for representation. My long-term objective is to shift the paradigm of AI model designs for computational pathology, such that it enables engineering-science developers to bridge the existing gaps from end-to-end viewpoint. My short-term objectives are: 1. Develop generalization measures in deep training for effective and eXplainable encoding of model parameters. We will design advanced metric measures for encode training parameters and consolidate the measures to design hyper-parameter optimization algorithms. 2. Develop knowledge aggregation models to benefit from multiple-source domain data in computational pathology that are numerically efficient, compact, privacy-preserved, and eXplainable. We will merge advances from federated-learning, knowledge distillation, and generalization measures to design aggregation models. 3. Develop automated searching algorithms for optimum design of deep neural network architectures for computational pathology. We will merge advances from generalization measure, knowledge aggregation from multi-source data for designing optimum cell topology for image representation. 4. Develop supervised multi-label contrastive learning loss-functions that are efficient, generalizable, and explainable for pathology image representation. We will design an end-to-end supervised classifiers by modeling multi-modal probabilistic distances for loss-landscape designs. We will evaluate each of our objectives across many experiments and compare our solutions to existing model-centric and data-centric approach designs in the literature as well as considering EDI factors. This comprehensive research program bridges the existing gap for theoretical advancements in deep learning designs for computational pathology. It creates a highly original paradigm shift in designing deep learning models and contributes to groundbreaking advances in both computer vision and medical imaging application designs in general. This could be considered as the stepping-stone in advanced deep-learning problems to derive decisive actions for better design of optimization algorithms and model architectures for image representation.