Experimental Design for Multi-Channel Imaging via Task-Driven Feature Selection

TL;DR

A new paradigm for experimental design that simultaneously optimizes the design and trains a machine-learning model to execute a user-specified image-analysis task to shorten acquisition time, reduce costs, and accelerate the deployment of imaging devices is presented.

Abstract

This paper presents a data-driven, task-specific paradigm for experimental design, to shorten acquisition time, reduce costs, and accelerate the deployment of imaging devices. Current approaches in experimental design focus on model-parameter estimation and require specification of a particular model, whereas in imaging, other tasks may drive the design. Furthermore, such approaches often lead to intractable optimization problems in real-world imaging applications. Here we present a new paradigm for experimental design that simultaneously optimizes the design (set of image channels) and trains a machine-learning model to execute a user-specified image-analysis task. The approach obtains data densely-sampled over the measurement space (many image channels) for a small number of acquisitions, then identifies a subset of channels of prespecified size that best supports the task. We propose a method: TADRED for TAsk-DRiven Experimental Design in imaging, to identify the most informative channel-subset whilst simultaneously training a network to execute the task given the subset. Experiments demonstrate the potential of TADRED in diverse imaging applications: several clinically-relevant tasks in magnetic resonance imaging; and remote sensing and physiological applications of hyperspectral imaging. Results show substantial improvement over classical experimental design, two recent application-specific methods within the new paradigm, and state-of-the-art approaches in supervised feature selection. We anticipate further applications of our approach. Code is available: https://github.com/sbb-gh/experimental-design-multichannel

Figures (8)

• Figure 1: a) An example of experimental design for imaging. In remote sensing hyperspectral imaging (see table \ref{['indian_pine_results:table']}), each observed wavelength $\textbf{d}^{i}$ is chosen by the experimenter. The outcome of each $\textbf{d}^{i}$ is a grayscale image - a channel of the resultant data $X_{D}$ (RGB has $3$ channels). b) The new paradigm for experimental design illustrated for qMRI. First, obtain image data $X_{\bar{D}}$ with a large number of $\bar{C}$ channels. Next, train a user-chosen task network, which drives design optimization to select $C < \bar{C}$ channels -- we propose TADRED for this. We consider three distinct example tasks in experiments.
• Figure 2: Downstream MRI metrics (see appendix \ref{['hcp_data_fitting_models:sec']}) estimated from the full set of channels/measurements on HCP data $\bar{C}=288$, and $\bar{C}$ from $C=18$ reconstructed measurements. Left: MSE for various metrics; Right: Qualitative comparison where arrows highlight closer agreement from TADRED's design with the gold standard than those from the best performing baseline.
• Figure 3: TADRED's structure. During training TADRED concurrently performs feature scoring, feature subsampling, and task execution. During training we progressively set the score to be sample-independent by setting $\alpha$ to $1$. We score features with $\bar{\textbf{s}}_{t} \in \mathbb{R}^{\bar{C}}$ and remove features with low score by setting corresponding values of the mask to $0$, in this example we removed feature 2.
• Figure 4: Analyzing the performance on different densely-sampled designs $\bar{D}$ where $|\bar{D}| = \bar{C}$ and $C = 14$. Settings follow table \ref{['parameter_estimate_VERDICT:table']}.
• Figure 5: Correlation coefficient between the measurements/features/channels of the data.