Problem Description

An AI Executive from a well-known medical imaging company approached us about guidance in developing a system to detect body regions from 3D medical images. A 3D medical scan is sliced into a sequence of 2D images. The sequence of images is then classified. The challenges that were faced by the client included what infrastructure to use, how to improve accuracy, how to invest in labeling and how to incorporate additional structure.

Reliancy Solution

Our Guidance stretched over multiple years and included:

• Leadership Support

  Roadmap development. Infrastructure assessments. Education of leaders on most applicable approaches.

• Accuracy Improvements

  Exploration of sequence models to gain accuracy over basic classifiers. Sequence models take into account all image slices in sequence, rather than looking at them independently.

• Optimizing Labeling Efforts

  Exploration on how to leverage active learning. Labeling in this domain is expensive. Active learning identifies which images would add most value if labeled.

• Exploiting Added Problem Structure

  Extending the basic classifier to incorporate other forms of data by leveraging a graph-based neural network.

Impact

Our guidance significantly affected the direction of product development.

• Assisted Client In Navigating A Challenging Problem

  Using our experience we helped the client avoid pitfalls. We helped them understand where the low hanging fruits are what next steps were most likely to lead to success.

• Time and Cost Reduction

  We helped our client understand how to most effectively label their data. Beyond that we provided corporate education both the the executives and the team members. By focusing our client on what's relevant we effecitvely helped them save time and cost.

• The Approach Gained FDA Approval

  We are proud that a product which received so much of our input ended up gaining FDA approval.