zea.agent

Agent subpackage for closing action-perception loop in ultrasound imaging.

The zea.agent subpackage provides tools and utilities for implementing cognitive ultrasound imaging via active perception. This module enables intelligent, adaptive transmit design, where acquisition decisions are informed by the current belief state about the imaged tissue.

Overview

Active perception in ultrasound involves iteratively:

1. Perceiving the current state of tissue from acquired measurements

2. Selecting transmit actions based on the current beliefs about the tissue state

3. Acquiring new data and looping back to the perception step

The zea.agent module provides the building blocks for implementing such perception-action loops.

Note

The functions currently available implement selection strategies for focused transmit actions. Development of action selection functions for more general transmit schemes is currently a work-in-progress.

Action Selection Strategies

Action selection strategies determine which transmits to fire next, given some belief about the tissue state.

See the following dropdown for a list of available action selection strategies:

Available strategies

• zea.agent.selection.GreedyEntropy: Selects lines that maximize entropy reduction.

• zea.agent.selection.UniformRandomLines: Randomly samples scan lines with uniform probability.

• zea.agent.selection.EquispacedLines: Selects equispaced lines that sweep across the image.

• zea.agent.selection.CovarianceSamplingLines: Models line-to-line correlation to select masks with highest entropy.

• zea.agent.selection.TaskBasedLines: Selects lines to maximize information gain with respect to a downstream task.

Basic Usage

>>> import zea
>>> import numpy as np

>>> agent = zea.agent.selection.GreedyEntropy(
...     n_actions=7,
...     n_possible_actions=112,
...     img_width=112,
...     img_height=112,
... )

>>> # (batch, samples, height, width)
>>> particles = np.random.rand(1, 10, 112, 112)
>>> lines, mask = agent.sample(particles)

Masks

The zea.agent.masks module provides utilities for converting action representations (e.g., selected line indices) to image-sized masks that can be applied to observations.

Example Notebooks

We provide example notebooks demonstrating perception-action loops in practice, as companions to recently published papers on the topic:

Patient-Adaptive Echocardiography

This tutorial implements a basic perception-action loop using diffusion models for perception-as-inference and greedy entropy minimization for action selection.

• Active perception for focused transmit steering

  • Uses GreedyEntropy to select informative scan lines

  • Demonstrates iterative belief refinement with sparse acquisitions

  • Visualizes the reconstruction process over multiple acquisition steps

Task-Based Transmit Beamforming

This tutorial implements a task-driven perception-action loop where acquisition decisions are optimized to gain information about a specific downstream task.

• Task-based transmit beamforming perception-action loop

  • Uses TaskBasedLines for task-aware line selection

  • Computes saliency maps via uncertainty propagation through downstream task models

  • Demonstrates how to integrate domain-specific measurement tasks

Modules

gumbel	Gumbel-Softmax trick implemented with the multi-backend keras.ops.
masks	Mask generation utilities.
selection	Action selection strategies