Specialized computer-aided diagnosis and disease characterization with a multi-focal ensemble of convolutional neural networks
Inventors
Madabhushi, Anant • Braman, Nathaniel • Maidment, Tristan • Chen, Yijiang
Assignees
Case Western Reserve University • US Department of Veterans Affairs
Publication Number
US-11817204-B2
Publication Date
2023-11-14
Expiration Date
2040-12-09
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
Embodiments discussed herein facilitate determination of whether lesions are benign or malignant. One example embodiment is a method, comprising: accessing medical imaging scan(s) that are each associated with distinct angle(s) and each comprise a segmented region of interest (ROI) of that medical imaging scan comprising a lesion associated with a first region and a second region; providing the first region(s) of the medical imaging scan(s) to trained first deep learning (DL) model(s) of an ensemble and the second region(s) of the medical imaging scan(s) to trained second DL model(s) of the ensemble; and receiving, from the ensemble of DL models, an indication of whether the lesion is a benign architectural distortion (AD) or a malignant AD.
Core Innovation
The invention facilitates determining whether lesions, specifically architectural distortions (AD) on medical imaging scans such as 3D Digital Breast Tomosynthesis (DBT), are benign or malignant by employing an ensemble of deep learning (DL) models trained on distinct regions and views of the lesions. Each medical imaging scan includes segmented regions of interest (ROI) containing lesions, and the ensemble inputs these regions from different angles and locations to respective trained DL models, receiving an indication of lesion classification as benign or malignant AD.
The problem addressed arises from the challenge of distinguishing benign architectural distortions such as radial scars (RS) from malignant AD, which share visual similarities and are difficult to differentiate using conventional methods. While 3D DBT offers improved sensitivity over 2D mammography, and deep learning methods have shown potential for breast cancer diagnosis, rare and complex confounders like AD remain problematic for general DL approaches due to limited and ambiguous training data. The invention solves this by constructing specialized ensembles of DL models trained on spatially and view-specific lesion samples with probabilistic weighting, improving diagnostic accuracy and reducing unnecessary biopsies.
Claims Coverage
The patent includes a set of independent claims directed to methods and apparatus involving ensembles of deep learning models trained and applied to medical imaging scans, with focus on spatially weighted lesion regions and multiple views.
Weighted region-based lesion analysis employing probabilistic spatial weighting
The method applies different probabilistic weights to lesion regions based on spatial positions within segmented ROIs of medical imaging scans, using probability distributions such as Gaussian distributions to weight intralesional and perilesional regions separately during training and inference.
Ensemble of convolutional neural networks specialized by view and region
An ensemble of deep learning models, including convolutional neural networks with volumetric kernels, is trained where each model specializes in a unique combination of lesion region (intralesional or perilesional) and imaging view (e.g., craniocaudal or mediolateral). The ensemble combines outputs from these independently trained models to classify lesions as benign or malignant.
Training using weighted samples from distinct lesion regions and views
Training involves extracting volumetric samples from the first and second lesion regions of medical imaging scans at distinct angles, applying associated weights based on probabilistic spatial information, and training respective DL models for each region-view combination. Model weightings within the ensemble are also determined to optimize performance.
Use of medical imaging scans from distinct angles in ensemble inputs
At least two medical imaging scans from distinct angles (e.g., craniocaudal and mediolateral) are accessed, each including segmented lesion ROIs. Different regions from these distinct angle scans are input to corresponding DL models within the ensemble, ensuring diverse data representation per model specializing in a view-region combination.
The independent claims focus on an ensemble deep learning approach to classify architectural distortion lesions by employing spatially weighted lesion regions from multiple imaging views. The ensemble consists of DL models, mainly convolutional neural networks with volumetric kernels, trained with weighted samples to improve discrimination between benign and malignant lesions.
Stated Advantages
Improved ability to distinguish benign from malignant architectural distortions, thereby reducing unnecessary biopsies without misclassifying malignant lesions.
Enhanced diagnostic performance by combining multiple DL models specialized on different lesion regions and imaging views, leveraging the unique variance of weak classifiers.
Provision of a probabilistic weighting scheme to better represent ambiguous lesion boundaries, improving model generalization in limited and complex datasets.
Potential for automated, computer-aided diagnosis that supplements visual assessment, aiding clinicians in handling rare and challenging confounders.
Documented Applications
Diagnostic classification of architectural distortions in 3D digital breast tomosynthesis medical imaging scans to determine whether lesions are benign or malignant.
Computer-aided diagnosis systems aimed at reducing unnecessary biopsies by improving accuracy in differentiating rare benign lesions such as radial scars from malignant lesions.
Training and deploying specialized deep learning models within medical imaging analysis pipelines, particularly for breast cancer detection using multiple imaging views and spatial lesion regions.
Use in personalized medicine devices or CADx systems to facilitate monitoring and treatment decisions based on classified lesion malignancy from medical images.
Interested in licensing this patent?