Deep Movement Primitives: Toward Breast Cancer Examination Robot

Oluwatoyin Sanni, Giorgio Bonvicini, Muhammad Arshad Khan, Pablo C. López-Custodio, Kiyanoush Nazari, Amir M. Ghalamzan E.AAAI 2022 · DOI

Introduction

Breast cancer remains the most prevalent cancer worldwide, making early detection through clinical breast palpation (BP) essential. However, human examinations are often limited by practitioner expertise and patient accessibility. While Autonomous Robotic Breast Palpation (ARBP) offers a path toward standardized care, it faces significant challenges in motion planning. Breasts vary significantly in geometry, and expert palpation requires complex, specific patterns such as spirals or wedges. Traditional "Learning from Demonstration" (LfD) techniques often struggle to directly map visual sensory data to these intricate manipulation paths.

Standard clinical breast palpation patterns: spiral, wedge, and vertical strip.

The deep-MP Framework

To address these challenges, researchers developed deep Movement Primitives (deep-MP). This novel framework bridges the gap between computer vision and trajectory generation by combining Probabilistic Movement Primitives (ProMP) with Deep Neural Networks.

Instead of learning a single, rigid path, the robot learns a distribution of trajectories via ProMP, providing the mathematical flexibility needed to reach various goals or "via points." The core innovation lies in using Deep Neural Networks—such as CNNs or PointNets—to process images of the workspace and output the specific weights ($\Theta$) that define the ProMP trajectory. To enhance precision, the authors proposed Residual deep-MP, where the network predicts the difference (residual) between a standard mean trajectory and the specific path required for a given visual scene.

Experimental Setup and Datasets

The system was implemented using a 7-DOF Franka Emika Panda robotic arm equipped with an Intel RealSense RGB-D camera and a high-resolution Xela tactile sensor.

Experimental setup featuring the Panda robotic arm, RGB-D camera, and Xela tactile sensor positioned over a breast phantom.

Due to the lack of existing datasets, the team created three custom datasets using a silicone breast phantom with implanted "lumps":

  • RTP-RGB/RGBD: Samples of "Reach-to-Palpate" actions using RGB and Depth data.
  • WPP (Wedges Palpation Path): Samples of the robot performing specific "wedge" palpation patterns across various phantom configurations.

Key Results and Insights

The deep-MP approach demonstrated superior accuracy and generalization capabilities:

  • High Precision: In the Wedges Palpation Path task, the CNN-based deep-MP model achieved a Mean Squared Error (MSE) of just 0.0072 $rad^2$, significantly outperforming baseline methods.
  • Effective Mapping: The "Residual deep-MP" approach proved highly effective for Reach-to-Palpate tasks, reducing Cartesian error to 24.6 mm in optimal regions.
  • Vision-to-Motion: CNNs processing 2D images generally outperformed PointNets processing 3D point clouds, suggesting that high-resolution spatial features in images are critical for this task.
  • Generalization: The model successfully adapted to unseen phantom configurations, proving it learned the underlying relationship between visual features (like nipple position or phantom curvature) and required joint space trajectories.

Limitations and Future Outlook

While a significant step forward, the study currently relies on a single type of breast phantom. Real-world applications will require handling the vast variability in human tissue stiffness and anatomy. Furthermore, while the robot included a tactile sensor, its data was not yet used for real-time path adjustment. Future work aims to integrate this tactile feedback for "closed-loop" palpation and conduct clinical validation to move the technology from the lab to the clinic.

Summary updated

BibTeX
@inproceedings{sanni2022deep,
  author = {Oluwatoyin Sanni and Giorgio Bonvicini and Muhammad Arshad Khan and Pablo C. López-Custodio and Kiyanoush Nazari and Amir M. Ghalamzan E.},
  title = {{Deep Movement Primitives: Toward Breast Cancer Examination Robot}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21472},
}
Figure