The Tuner workflow supports developers of image-segmentation algorithms in systematically finding “good” parameter settings, replacing tedious manual trial-and-error with a principled exploration of the multi-dimensional parameter space. Given a segmentation algorithm controlled by several tunable parameters, a ground-truth segmentation, and one or two numerical quality measures, the workflow follows a two-stage strategy: (1) sparsely sample the parameter space and fit a statistical model that estimates the response (segmentation goodness) together with its uncertainty; and (2) interactively explore and iteratively refine the response surface to locate high-quality, stable parameter regions.
In the first stage, the system places an initial set of sample points (parameter configurations) across the full parameter space using Latin Hypercube or random sampling. Each sample point is passed to the external segmentation algorithm (run non-interactively as a batch process) to produce a segmented image. Each segmentation is then compared against the ground-truth image using the chosen quality metrics (e.g. Dice coefficient and recovery error, or Precision and Recall), yielding scalar response values per sample point. From these design points a Gaussian process (DACE) model is fitted, interpolating a continuous response surface and providing a per-point uncertainty (standard error) and an expected-improvement (gain) estimate.
The estimated response surface is arranged as a HyperSlice matrix of orthogonal 2D slices centred on a focus point, and rendered as a Response View offering three interchangeable plot types (response value, estimation error/uncertainty, and expected gain). Two further linked views are produced: a Pareto Panel scatterplot showing the trade-off between two quality measures across sampled points, and response histograms showing the estimated distribution of each response dimension.
In the second stage, the analyst enters an iterative exploration and refinement loop. Starting from a favourable trade-off point selected in the Pareto Panel, the analyst navigates the slices (clicking, dragging, sliders, zooming, colormap compression) to locate high-response regions and to judge parameter sensitivity from how strongly the slices change. The analyst assesses whether a stable, well-sampled region meeting the quality constraints has been found. If not, the analyst marks a hyperbox region of interest — guided by the expected-gain plots — and adds new sample points there; these are segmented, evaluated against ground truth, the Gaussian process model is automatically rebuilt on the augmented design points, and all views are re-rendered. The cycle repeats until a satisfactory parameter region is identified.
Finally, in a knowledge synthesis step, the analyst selects the final parameter setting or stable range from the identified region (and, when tuning across several datasets, takes the intersection of the valid ranges), together with an understanding of how sensitive the segmentation is to each parameter.