Arizona State University

Mayo Foundation for Medical Education and Research

Arizona Board of Regents, a Body Corporate

Jianming Liang

Nima Tajbakhsh

Zongwei Zhou

Mohammad Reza Hosseinzadeh Taher

Fatemeh Haghighi

Suryakanth R Gurudu

Md Mahfuzur Rahman Siddiquee

Hong Wu

Ruibin Feng

Jae Shin

Wenzhe Xue

Jiaxuan Pang

Nahid Ul Islam

Vatsal Sodha

Marek Belohlavek

Eileen M McMahon

Wei Liu

Christopher B Kendall

Robert T Hurst

Changching Chi

Xiangjun Zhu

Haripriya Sharma

Kun Deng

DongAo Ma

Jaeyul Shin

Yu An

Shiv Gehlot

Shivam Bajpai

Esra Dincer

Diksha Goyal

Michael Gotway

A self-supervised learning framework for empowering instance discrimination in medical imaging using Context-Aware instance Discrimination (CAiD), in which the trained deep models are then utilized for the processing of medical imaging. An exemplary system receives a plurality of medical images; trains a self-supervised learning framework to increasing instance discrimination for medical imaging using a Context-Aware instance Discrimination (CAiD) model using the received plurality of medical images; generates multiple cropped image samples and augments samples using image distortion; applies instance discrimination learning a mapping back to a corresponding original image; reconstructs the cropped image samples and applies an auxiliary context-aware learning loss operation; and generates as output, a pre-trained CAiD model based on the application of both (i) the instance discrimination learning and (ii) the auxiliary context-aware learning loss operation.

Systems, methods, and apparatuses for implementing a self-supervised learning framework for empowering instance discrimination in medical imaging using context-aware instance discrimination (CAiD)

Described herein are means for the generation of Transferable Visual Word (TransVW) models through self-supervised learning in the absence of manual labeling, in which the trained TransVW models are then utilized for the processing of medical imaging. For instance, an exemplary system is specially configured to perform self-supervised learning for an AI model in the absence of manually labeled input, by performing the following operations: receiving medical images as input; performing a self-discovery operation of anatomical patterns by building a set of the anatomical patterns from the medical images received at the system, performing a self-classification operation of the anatomical patterns; performing a self-restoration operation of the anatomical patterns within cropped and transformed 2D patches or 3D cubes derived from the medical images received at the system by recovering original anatomical patterns to learn different sets of visual representation; and providing a semantics-enriched pre-trained AI model having a trained encoder-decoder structure with skip connections in between based on the performance of the self-discovery operation, the self-classification operation, and the self-restoration operation. Other related embodiments are disclosed.

Systems, methods, and apparatuses for implementing transferable visual words by exploiting the semantics of anatomical patterns for self-supervised learning

Described herein are means for implementing contrastive learning via reconstruction within a self-supervised learning framework, in which the trained deep models are then utilized for the processing of medical imaging. For instance, an exemplary system is specially configured for performing a random cropping operation to crop a 3D cube from each of a plurality of medical images received at the system as input; performing a resize operation of the cropped 3D cubes; performing an image reconstruction operation of the resized and cropped 3D cubes to predict the whole image represented by the original medical images received; and generating a reconstructed image which is analyzed for reconstruction loss against the original image representing a known ground truth image to the reconstruction loss function. Other related embodiments are disclosed.

Systems, methods, and apparatuses for implementing contrastive learning via reconstruction within a self-supervised learning framework

Described herein are systems, methods, and apparatuses for implementing self-supervised domain-adaptive pre-training via a transformer for use with medical image classification in the context of medical image analysis. An exemplary system includes means for receiving a first set of training data having non-medical photographic images; receiving a second set of training data with medical images; pre-training an AI model on the first set of training data with the non-medical photographic images; performing domain-adaptive pre-training of the AI model via self-supervised learning operations using the second set of training data having the medical images; generating a trained domain-adapted AI model by fine-tuning the AI model against the targeted medical diagnosis task using the second set of training data having the medical images; outputting the trained domain-adapted AI model; and executing the trained domain-adapted AI model to generate a predicted medical diagnosis from an input image not present within the training data.

Systems, methods, and apparatuses for implementing self-supervised domain-adaptive pre-training via a transformer for use with medical image classification

A Discriminative, Restorative, and Adversarial (DiRA) learning framework for self-supervised medical image analysis is described. For instance, a pre-trained DiRA framework may be applied to diagnosis and detection of new medical images which form no part of the training data. The exemplary DiRA framework includes means for receiving training data having medical images therein and applying discriminative learning, restorative learning, and adversarial learning via the DiRA framework by cropping patches from the medical images; inputting the cropped patches to the discriminative and restorative learning branches to generate discriminative latent features and synthesized images from each; and applying adversarial learning by executing an adversarial discriminator to perform a min-max function for distinguishing the synthesized restorative image from real medical images. The pre-trained model of the DiRA framework is then provided as output for use in generating predictions of disease within medical images.

Systems, methods, and apparatuses for implementing discriminative, restorative, and adversarial (DiRA) learning for self-supervised medical image analysis

Embodiments described herein include systems for implementing annotation-efficient deep learning in computer-aided diagnosis. Exemplary embodiments include systems having a processor and a memory specially configured with instructions for learning annotation-efficient deep learning from non-labeled medical images to generate a trained deep-learning model by applying a multi-phase model training process via specially configured instructions for pre-training a model by executing a one-time learning procedure using an initial annotated image dataset; iteratively re-training the model by executing a fine-tuning learning procedure using newly available annotated images without re-using any images from the initial annotated image dataset; selecting a plurality of most representative samples related to images of the initial annotated image dataset and the newly available annotated images by executing an active selection procedure based on the which of a collection of un-annotated images exhibit either a greatest uncertainty or a greatest entropy; extracting generic image features; updating the model using the generic image features extracted; and outputting the model as the trained deep-learning model for use in analyzing a patient medical image. Other related embodiments are disclosed.

Systems, methods, and apparatuses for implementing advancements towards annotation efficient deep learning in computer-aided diagnosis

Described herein are means for generating pre-trained models for nnU-Net through the use of improved transfer learning techniques, in which the pre-trained models are then utilized for the processing of medical imaging. According to a particular embodiment, there is a system specially configured for segmenting medical images, in which such a system includes: a memory to store instructions; a processor to execute the instructions stored in the memory; wherein the system is specially configured to: execute instructions via the processor for executing a pre-trained model from Models Genesis within a nnU-Net framework; execute instructions via the processor for learning generic anatomical patterns within the executing Models Genesis through self-supervised learning; execute instructions via the processor for transforming an original image using distortion and cutout-based methods; execute instructions via the processor for learning the reconstruction of the original image from the transformed image using an encoder-decoder architecture of the nnU-Net framework to identify the generic anatomical representation from the transformed image by recovering the original image; and wherein architecture determined by the nnU-Net framework is utilized with Models Genesis and is trained to minimize the L2 distance between the prediction and ground truth. Other related embodiments are disclosed.

Systems, methods, and apparatuses for generating pre-trained models for nnU-net through the use of improved transfer learning techniques

Described herein are means for the generation of semantic genesis models through self-supervised learning in the absence of manual labeling, in which the trained semantic genesis models are then utilized for the processing of medical imaging. For instance, an exemplary system is specially configured with means for performing a self-discovery operation which crops 2D patches or crops 3D cubes from similar patient scans received at the system as input; means for transforming each anatomical pattern represented within the cropped 2D patches or the cropped 3D cubes to generate transformed 2D anatomical patterns or transformed 3D anatomical patterns; means for performing a self-classification operation of the transformed anatomical patterns by formulating a C-way multi-class classification task for representation learning; means for performing a self-restoration operation by recovering original anatomical patterns from the transformed 2D patches or transformed 3D cubes having transformed anatomical patterns embedded therein to learn different sets of visual representation; and means for providing a semantics-enriched pre-trained AI model having a trained encoder-decoder structure with skip connections in between based on the performance of the self-discovery operation, the self-classification operation, and the self-restoration operation. Other related embodiments are disclosed.

Systems, methods, and apparatuses for the use of transferable visual words for AI models through self-supervised learning in the absence of manual labeling for the processing of medical imaging

Described herein are means for generating source models for transfer learning to application specific models used in the processing of medical imaging. In some embodiments, the method comprises: identifying a group of training samples, wherein each training sample in the group of training samples includes an image; for each training sample in the group of training samples: identifying an original patch of the image corresponding to the training sample; identifying one or more transformations to be applied to the original patch; generating a transformed patch by applying the one or more transformations to the identified patch; and training an encoder-decoder network using a group of transformed patches corresponding to the group of training samples, wherein the encoder-decoder network is trained to generate an approximation of the original patch from a corresponding transformed patch, and wherein the encoder-decoder network is trained to minimize a loss function that indicates a difference between the generated approximation of the original patch and the original patch. The source models significantly enhance the transfer learning performance for many medical imaging tasks including, but not limited to, disease/organ detection, classification, and segmentation. Other related embodiments are disclosed.

Systems, methods, and apparatuses for the generation of source models for transfer learning to application specific models used in the processing of medical imaging

Described herein are means for systematically determining an optimal approach for the computer-aided diagnosis of a pulmonary embolism, in the context of processing medical imaging. According to a particular embodiment, there is a system specially configured for diagnosing a Pulmonary Embolism (PE) within new medical images which form no part of the dataset upon which the AI model was trained. Such a system executes operations for receiving a plurality of medical images and processing the plurality of medical images by executing an image-level classification algorithm to determine the presence or absence of a Pulmonary Embolism (PE) within each image via operations including: pre-training an AI model through supervised learning to identify ground truth; fine-tuning the pre-trained AI model specifically for PE diagnosis to generate a pre-trained PE diagnosis and detection AI model; wherein the pre-trained AI model is based on a modified CNN architecture having introduced therein a squeeze and excitation (SE) block enabling the CNN architecture to extract informative features from the plurality of medical images by fusing spatial and channel-wise information; applying the pre-trained PE diagnosis and detection AI model to new medical images to render a prediction as to the presence or absence of the Pulmonary Embolism within the new medical images; and outputting the prediction as a PE diagnosis for a medical patient.

Systems, methods, and apparatuses for systematically determining an optimal approach for the computer-aided diagnosis of a pulmonary embolism

Described herein are means for implementing fixed-point image-to-image translation using improved Generative Adversarial Networks (GANs). For instance, an exemplary system is specially configured for implementing a new framework, called a Fixed-Point GAN, which improves upon prior known methodologies by enhancing the quality of the images generated through global, local, and identity transformation. The Fixed-Point GAN as introduced and described herein, improves many applications dependant on image-to-image translation, including those in the field of medical image processing for the purposes of disease detection and localization. Other related embodiments are disclosed.

Systems, methods, and apparatuses for implementing fixed-point image-to-image translation using improved generative adversarial networks (GANs)

Described herein are systems, methods, and apparatuses for actively and continually fine-tuning convolutional neural networks to reduce annotation requirements, in which the trained networks are then utilized in the context of medical imaging. The success of convolutional neural networks (CNNs) in computer vision is largely attributable to the availability of massive annotated datasets, such as ImageNet and Places. However, it is tedious, laborious, and time consuming to create large annotated datasets, and demands costly, specialty-oriented skills. A novel method to naturally integrate active learning and transfer learning (fine-tuning) into a single framework is presented to dramatically reduce annotation cost, starting with a pre-trained CNN to seek 'worthy' samples for annotation and gradually enhances the (fine-tuned) CNN via continual fine-tuning. The described method was evaluated using three distinct medical imaging applications, demonstrating that it can reduce annotation efforts by at least half compared with random selection.

Systems, methods, and apparatuses for actively and continually fine-tuning convolutional neural networks to reduce annotation requirements

Systems for selecting candidates for labelling and use in training a convolutional neural network (CNN) are provided, the systems comprising: a memory device; and at least one hardware processor configured to: receive a plurality of input candidates, wherein each candidate includes a plurality of identically labelled patches; and for each of the plurality of candidates: determine a plurality of probabilities, each of the plurality of probabilities being a probability that a unique patch of the plurality of identically labelled patches of the candidate corresponds to a label using a pre-trained CNN; identify a subset of candidates of the plurality of input candidates, wherein the subset does not include all of the plurality of candidates, based on the determined probabilities; query an external source to label the subset of candidates to produce labelled candidates; and train the pre-trained CNN using the labelled candidates.

Systems, methods, and/or media, for selecting candidates for annotation for use in training a classifier

Medical image segmentation using interactive refinement, in which the trained deep models are then utilized for the processing of medical imaging are described. Operating a two-step deep learning training framework including receiving original input images at the deep learning training framework; generating an initial prediction image specifying image segmentation by base segmentation model; receiving user input guidance signals; routing each of (i) the original input images, (ii) the initial prediction image, and (iii) the user input guidance signals to an InterCNN; generating a refined prediction image specifying refined image segmentation by processing each of the (i) the original input images, (ii) the initial prediction image, and (iii) the user input guidance signals through the InterCNN to render the refined prediction image incorporating the user input guidance signals; and outputting a refined segmentation mask to the deep learning training framework as a guidance signal.

Systems, methods, and apparatuses for implementing medical image segmentation using interactive refinement

Described herein are means for generation of self-taught generic models, named Models Genesis, without requiring any manual labeling, in which the Models Genesis are then utilized for the processing of medical imaging. For instance, an exemplary system is specially configured for learning general-purpose image representations by recovering original sub-volumes of 3D input images from transformed 3D images. Such a system operates by cropping a sub-volume from each 3D input image; performing image transformations upon each of the sub-volumes cropped from the 3D input images to generate transformed sub-volumes; and training an encoder-decoder architecture with skip connections to learn a common image representation by restoring the original sub-volumes cropped from the 3D input images from the transformed sub-volumes generated via the image transformations. A pre-trained 3D generic model is thus provided, based on the trained encoder-decoder architecture having learned the common image representation which is capable of identifying anatomical patterns in never before seen 3D medical images having no labeling and no annotation. More importantly, the pre-trained generic models lead to improved performance in multiple target tasks, effective across diseases, organs, datasets, and modalities.

Growing community of inventors

Scottsdale, AZ, United States of America

Jianming Liang