Authors:

Byung Wook Kim

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Ultrasound scanning is the most widely used medical imaging modality for various diagnostic procedures and its utilization has increased significantly in resource limited settings. This is due to its affordable cost, non- radiative nature, and ability to provide real-time diagnostic feedback. However, it is a highly expert-dependent modality, which often leads to inconsistent results, erroneous diagnoses, and inaccessibility in rural areas where expert sonographers are available in limited capacity. To address this issue, Autonomous Robotic Ultrasound Systems (A-RUS) implementing model-based and model-free methods for robotic probe motion were introduced. Yet, these existing systems failed to incorporate expert knowledge and technique in their framework leading to their limited clinical adaptability in diverse physiological procedures. In this research, we propose a learning-from-expert demonstration framework to train the RUS that emulates the training process of novice sonographers under expert supervision. First, the RUS is trained using a dataset comprising probe motions and corresponding ultrasound images acquired during the expert's demonstration. Then, the model aims to learn the expert's probing technique using behavioral cloning framework and to predict the next robotic probe movement that an expert sonographer would have performed. The expected outcome of this framework is the prior motion model that can be executed on a RUS for imitating the expert's decision-making skills. We will conduct rigorous validation studies by comparing the proposed RUS for mimicking the skills of different ultrasound experts at IU School of Medicine. Later, this model can be subsequently trained using reinforcement learning policies for efficient intra-patient robotic ultrasound procedures.

Source:

Purdue University / 2023

Topics:

No topics listed

Co-authors:

Byung Wook Kim