Efficient Image Analysis for Low Quality Medical Images

Abstract

Medical images captured with low-cost devices have some similar low quality characteristics such as a small field of view (FOV) compared to the overall size of the imaged area, low or degraded spatial resolution, and repetitive background with sparse features. Different from images of general scenarios with many distinctive features, these low quality medical images need more targeted approaches for analysis. This study focuses on the 2D medical images captured from different low-cost imaging devices. The medical images and videos are analyzed with two types of feature learning: one is based on the unsupervised dimension reduction and another one is based on the supervised neural network. In the first type, the coarse-to-fine frameworks for different downstream image analysis tasks are proposed. As the coarse step, the dimension reduction learns the embedding of input images in a low dimensional space which allows more efficient and reliable computation. It provides an efficient initialization for different methods in the following fine step, which narrows their optimization or search domain. The developed coarse-to-fine frameworks outperform classic methods considerably in template matching, mosaicking and localization tasks over low quality medical images and videos, including synthetic data, phantom data and in vivo data. It is the first proposed two-stage framework leveraging dimension reduction for efficient analysis of low-quality images, which can also be generalized into different downstream tasks by modifying the fine stage and achieve the state-of-the-art (SOTA) performance.In the second type, the deep learning based analysis is studied in the real-time eye tracking by localizing small FOV retina frames with large distortion. After backbone of the convolution neural network, the feature map of the input frame is designed as the kernel to perform convolution on the feature map of the full retina in the following convolution layers. To achieve a more robust performance under the frame distortion, a Kalman filter is used to embed the deep learning in the state transition model and the feature-point based template matching as the measurement. Combining neural network with feature point matching to reduce the noise in two process, the motion distortion is eliminated in the tracking result which outperforms relying on neural network only. A robotic platform is built to collect annotated dataset for supervised training of the network. Considering the overfitting to the noisy labels during training, the approach of learning with noisy labeling is explored. The proposed framework combines the idea of small loss selection and noise label correction, which learns network parameters and reassigns ground truth labels iteratively. The proposed method achieves the SOTA performance on both synthesized and real noise labelled dataset. It is independent to the backbone network structure and can be directly used in the training of different models by adding a Siamese network. The Siamese network is removed after training, the original model then can be used for test.