Optimizing Deep Learning Models for Medical Imaging and Genomics

摘要

Deep Learning is transforming the medicine field by enabling faster diagnosis, personalized treatments and wider access to healthcare. Deep Learning based frameworks such as U-shaped Convolutional Neural (U-Net), Residual Neural Network (ResNet) and Densely Connected Networks (DenseNet) detect neurological disorders and different kinds of cancers at their early stages with at-most accuracy from the PET/CT and MR images. Beyond imaging, deep learning models are widely used for speeding up drug discovery, robotic-assisted surgical procedures, wearable devices and remote care via telemedicine by revealing genetic markers. However, these models are continuously refined and optimized for accuracy enhancements, as they are dealing with human lives. This chapter covers the wide range of optimizations incorporated during training and pre-processing of data. Transfer learning, domain adaptation and ensemble learning enhance model performance. Also, it discusses the practicality of implementation, regressive concerns, computational efficiency and deployment challenges.