Obviousness

The obviousness of US Patent 11129591 under 35 U.S.C. § 103 can be analyzed by combining existing prior art and the general knowledge of a person having ordinary skill in the art (POSA) in echocardiographic image analysis and machine learning. The patent itself provides significant insights into the known art and the problems it aims to address.

Prior Art References Considered:

1. Admitted Prior Art (APA): The patent's background explicitly states, "Some existing echocardiographic systems may be configured to provide feedback regarding general properties of captured images." This establishes that general echocardiographic image quality assessment was known prior to the invention.
2. General Clinical Knowledge (GCK) of Echocardiography: The patent describes the well-established practice of using "standard 2D echocardiographic views" (e.g., AP2, AP3, AP4, PSAX A, PSAX PM) for specific diagnostic purposes and "quantified clinical measurement of anatomical features," such as determining ejection fraction using images from AP2 and AP4 views via the 2D Method of Simpson. This knowledge also includes the understanding that "multiple views may be required in order to perform certain quantified clinical measurement of anatomical features."
3. Simonyan et al. (2015) - VGG Architecture: The patent explicitly references "Simonyan, K., Zisserman, A.: Very Deep Convolutional Networks for Large-Scale Image Recognition. International Conference on Learning Representations (ICRL) pp. 1-14 (2015)" for the VGG architecture, describing convolutional layers with kernels, max-pooling layers, and the doubling of kernel numbers in deeper layers. This paper teaches advanced convolutional neural network (CNN) architectures suitable for image analysis.
4. Kingma & Ba (2015) - Adam Optimizer: The patent also references "Kingma, D. P., Ba, J. L.: Adam: a Method for Stochastic Optimization, International Conference on Learning Representations 2015 pp. 1-15 (2015)" for the Adam optimizer. This paper teaches a widely adopted stochastic gradient-based optimization algorithm for training neural networks.
5. General Machine Learning Practice (GMLP): By 2016, the effective filing date, supervised machine learning, particularly with neural networks, was well-understood. This includes the practice of training neural networks using labeled datasets, where domain-specific tasks often require labels provided by human experts to serve as ground truth. The patent's training process explicitly uses "expert quality assessment values provided by an expert echocardiographer."

---

Obviousness Analysis for Claims related to Echocardiographic Image Analysis (Independent Claims 1, 10, 19, 39):

These claims generally cover a computer-implemented system/method for receiving echocardiographic images, associating them with a view category, determining a view category specific quality assessment value (representing suitability for quantified clinical measurement), and then associating/producing signals for this value, repeated for different view categories.

Combination: Admitted Prior Art (APA) + General Clinical Knowledge (GCK) + Simonyan et al. (VGG) + Kingma & Ba (Adam) + General Machine Learning Practice (GMLP).

Motivation for a Person Having Ordinary Skill in the Art (POSA) to Combine:

1. Recognized Deficiency in Prior Art: The patent explicitly identifies a problem with existing systems: "Some existing echocardiographic systems may be configured to provide feedback regarding general properties of captured images. However, this feedback may not assist echocardiographers in capturing high quality echocardiographic images for use in subsequent quantified clinical measurement of anatomical features." A POSA would clearly recognize the need to improve this general feedback to provide more useful information for clinical diagnosis and measurement.
2. Clinical Need for Specificity (GCK): A POSA in echocardiography is well aware that different "standard 2D echocardiographic views" are critical for assessing specific anatomical features and performing particular "quantified clinical measurement of anatomical features" (e.g., AP2 and AP4 for ejection fraction). Therefore, to address the identified deficiency and provide truly useful feedback for clinical measurement, it would be an obvious step to make the quality assessment specific to each view category.
3. Availability of Powerful Tools (Simonyan et al. & Kingma & Ba): By 2015, advanced neural network architectures (like the VGG architecture described by Simonyan et al.) and efficient training optimizers (like Adam described by Kingma & Ba) were widely known and proven effective for complex image analysis tasks, including classification and regression. A POSA seeking to develop a more sophisticated and automated image quality assessment system would naturally turn to these powerful machine learning tools.
4. Routine Application of Supervised Learning (GMLP): To develop an automated system that provides view-category-specific quality assessments reflecting suitability for clinical measurement, a POSA would routinely employ supervised learning. This involves training a neural network with a dataset of echocardiographic images labeled with objective quality scores. For a medical task, these labels would ideally come from "expert echocardiographers."
5. Predictable Outcome: Combining these known elements to solve the stated problem leads to a predictable outcome. Applying established neural network techniques (Simonyan et al., Kingma & Ba) to generate view-category specific quality assessment values for echocardiographic images (GCK), where the general quality assessment was previously known (APA), would predictably result in more precise and clinically relevant feedback, thereby overcoming the limitations of general quality assessment. The specific implementation of a "view category specific image assessment neural network" is a direct and obvious application of these combined principles.

---

Obviousness Analysis for Claims related to Training Neural Networks (Independent Claims 20, 29, 38, 46):

These claims generally cover a computer-implemented system/method for receiving echocardiographic training images associated with view categories, receiving expert quality assessment values (representing suitability for quantified clinical measurement), and training neural networks using this data, with at least a portion of each neural network associated with a specific view category.

Combination: Simonyan et al. (VGG) + Kingma & Ba (Adam) + General Clinical Knowledge (GCK) + General Machine Learning Practice (GMLP).

Motivation for a POSA to Combine:

1. Standard Neural Network Training (Simonyan et al. & Kingma & Ba): The fundamental principles of designing and training deep neural networks, including specific architectures (e.g., VGG-like convolutional and pooling layers) and optimization methods (e.g., Adam optimizer, mean absolute error loss function, stochastic gradient-based optimization), were well-established and taught by Simonyan et al. and Kingma & Ba prior to the patent's priority date. A POSA would routinely utilize these methods to train an image analysis model.
2. Necessity of Labeled Data (GMLP): Supervised learning, which is central to training neural networks for specific tasks, inherently requires labeled training data. The patent's method of training "to minimize the absolute difference between normalized quality assessment values assigned by a trained echocardiographer to training images" explicitly demonstrates supervised learning.
3. Source of Medical Labels - Experts: For a subjective and critical medical diagnostic task like echocardiographic image quality assessment, particularly concerning "suitability for quantified clinical measurement," the most reliable and authoritative labels would come from "expert echocardiographers." It is common practice in medical imaging AI to rely on human experts for ground truth generation.
4. Specificity of Labels and Architecture for Clinical Relevance (GCK): Given the clinical need for view-category specific quality assessments to guide quantified clinical measurements (as discussed under GCK), it would be obvious that the "expert quality assessment values" used for training must likewise be view-category specific and reflect suitability for these measurements. Furthermore, designing a "multi-stream network" with "shared layers" for common feature extraction and "view-specific layers" for specialized assessment per category, as described in the patent, is an obvious adaptation of known neural network architectures (e.g., in multi-task learning or transfer learning) to efficiently learn these nuanced, view-specific criteria. This approach allows the network to learn both general image features and specific features relevant to each view category.

---

Conclusion:

The claimed inventions in US Patent 11129591 would have been obvious to a person having ordinary skill in the art by the effective filing date. The patent addresses a recognized deficiency in existing echocardiographic systems (general image quality feedback is insufficient for quantified clinical measurements) by applying well-known machine learning techniques (neural networks, supervised learning with expert data, multi-task architectures) to provide view-category specific quality assessments. The motivation for this combination is explicitly outlined in the patent's own background and problem statement, driving a POSA to leverage readily available technological solutions to meet a clear clinical need.