CaptION Health Inc v. University Of British Columbia

1. Case Identification

• Case #: IPR2025-01066
• Patent #: 11,129,591
• Petitioner(s): Caption Health, Inc.
• Patent Owner(s): University of British Columbia
• Challenged Claims: 1-20

2. Patent Overview

• Title: Echocardiographic Image Analysis
• Brief Description: The ’591 patent discloses a computer-implemented system for assessing the quality of echocardiographic images. The system first associates an acquired image with one of a plurality of predetermined "view categories" and then uses a view-category-specific neural network to determine a quality assessment value for that image.

3. Grounds for Unpatentability

Ground 1: Obviousness over Krishnan and Lee - Claims 1-5 and 15-19 are obvious over Krishnan in view of Lee.

• Prior Art Relied Upon: Krishnan (Application # 2005/0251013) and Lee (Application # 2016/0247034).
• Core Argument for this Ground:
  • Prior Art Mapping: Petitioner argued that Krishnan discloses the core workflow of claim 1: a system that receives an echocardiographic image, associates it with a view category, and determines a quality assessment value. Krishnan teaches using a "bank of classifiers," which can be neural networks, to perform this assessment. However, during prosecution, the claims were amended to add the specific steps of (1) determining that a particular set of assessment parameters (i.e., a specific neural network) is associated with the image's view category, and (2) in response, inputting the image into that selected neural network. Petitioner contended that Lee explicitly teaches these missing steps. Lee describes a system that first classifies an image into a category and then selects a corresponding category-specific quality classifier from a plurality of stored classifiers to evaluate the image.
  • Motivation to Combine: Petitioner asserted that a Person of Ordinary Skill in the Art (POSITA) would combine Krishnan and Lee to improve the accuracy of quality assessments. Lee teaches that images in different categories should be evaluated using different assessment parameters, which Petitioner argued is common sense for echocardiography where different views highlight different anatomical structures. A POSITA would have been motivated to apply Lee’s method of selecting an appropriate category-specific classifier to Krishnan’s "bank of classifiers" to yield more reliable, view-specific quality scores.
  • Expectation of Success: Petitioner argued a POSITA would have had a high expectation of success, as the combination involves applying a known technique (Lee's classifier selection) to a known system (Krishnan's image analysis framework) to achieve a predictable improvement in performance.

Ground 2: Obviousness over Krishnan-Lee and Pagoulatos - Claims 7-9 and 11-13 are obvious over Krishnan and Lee in further view of Pagoulatos.

• Prior Art Relied Upon: Krishnan (Application # 2005/0251013), Lee (Application # 2016/0247034), and Pagoulatos (Application # 2017/0262982).
• Core Argument for this Ground:
  • Prior Art Mapping: This ground addresses claims directed to training the neural networks. Petitioner argued that the combination of Krishnan and Lee establishes a system with multiple, view-category-specific neural network classifiers. Pagoulatos was introduced to supply the detailed training methods recited in the claims. Pagoulatos explicitly describes training neural networks for ultrasound image analysis using a labeled training set of images with known characteristics (including various views and qualities) and using those images as inputs with their associated quality assessments as desired outputs.
  • Motivation to Combine: Petitioner argued that while Krishnan generally discloses a "learning engine" for building classifiers, Pagoulatos provides specific, conventional, and detailed techniques for training neural networks for the exact purpose of assessing echocardiographic images. A POSITA implementing the Krishnan-Lee system would have been motivated to use the established and advantageous training methods described in Pagoulatos to create the required view-specific classifiers. This combination represents using a known technique (Pagoulatos) with a known device (Krishnan-Lee) for its intended purpose.
  • Expectation of Success: Petitioner asserted a high expectation of success because Pagoulatos describes its method as "a common method of training artificial neural networks," making it a predictable and straightforward implementation.

Ground 3: Obviousness over Krishnan-Lee and Chen - Claims 6 and 20 are obvious over Krishnan and Lee in further view of Chen.

• Prior Art Relied Upon: Krishnan (Application # 2005/0251013), Lee (Application # 2016/0247034), and Chen (a 2015 conference paper).
• Core Argument for this Ground:
  • Prior Art Mapping: This ground targets claims requiring that the sets of assessment parameters for each neural network include both a "set of common assessment parameters" and a "set of view category specific assessment parameters." Petitioner asserted that the Krishnan-Lee combination teaches the use of separate, view-specific neural networks. Chen was added to teach the claimed multi-stream architecture. Chen discloses a joint learning model for ultrasound plane detection that uses a multi-stream convolutional neural network with shared layers (trained jointly across all tasks) and subsequent task-specific layers (trained individually for each task), corresponding directly to the claimed common and unique parameters.
  • Motivation to Combine: Petitioner contended that Chen explicitly teaches the advantages of this multi-task learning architecture, namely that it addresses insufficient training data and reduces overfitting—the same rationale later presented in the ’591 patent. A POSITA would have been motivated to structure the "bank of classifiers" from Krishnan-Lee according to Chen's architecture to gain these well-documented benefits of more efficient and robust training.
  • Expectation of Success: Petitioner argued for a high expectation of success, as the combination merely involves implementing a specific, known neural network architecture (multi-task learning) for which Krishnan’s system is agnostic and which Chen teaches is broadly applicable to anatomical structure detection problems.
• Additional Grounds: Petitioner asserted an additional obviousness challenge (Ground D for claims 10 and 14) based on the combination of Krishnan-Lee-Pagoulatos in further view of Chen, applying the multi-stream network architecture of Chen to the training methods taught by Pagoulatos.

4. Relief Requested

• Petitioner requests institution of an inter partes review and cancellation of claims 1-20 of Patent 11,129,591 as unpatentable.