International Business Machines Corporation

Merative US L.p.

Chun Lok Wong

Mehdi Moradi

Satyananda Kashyap

Tanveer Fathima Syeda-Mahmood

Hongzhi Wang

Alexandros Karargyris

Joy Tzung-yu Wu

Yaniv Gur

Anup Pillai

Ashutosh Jadhav

Hui Tang

Vaishnavi Subramanian

Joy Wu

Methods and systems for performing transfer learning with basis scaling and pruning. One method includes obtaining a pre-trained deep convolutional neural network (DCNN), decomposing each weight matrix of the DCNN, and decomposing each convolutional layer by applying the respective decomposed weight matrix to the convolution layer to form a first layer which comprises the left matrix for convolution, and a second layer which comprises the right matrix for convolution. The method also includes providing a basis-scaling convolutional layer having a weight matrix that is derived by a function of singular values and the right singular vectors and training the basis scaling factors of the basis-scaling convolutional layers.

Transfer learning with basis scaling and pruning

A computer-implemented method, a computer program product, and a computer system for introducing channel-scaling layers in a deep neural network. A computer receives a pre-trained deep neural network including convolutional layers followed by respective ones of activation layers, adds channel-scaling layers after the respective ones of the activation layers, where each of the channel-scaling layers includes scaling weights. The computer trains the scaling weights in the channel-scale layers. The computer removes, in the convolutional layers, channels whose corresponding scaling weights are lower than a predetermined threshold. The computer removes the channel-scaling layers. In response to determining that at least one convergence criterion is met, the computer provides a finally trained deep neural network.

Channel scaling: a scale-and-select approach for selective transfer learning

Disease detection from medical images is provided. In various embodiments, a medical image of a patient is read. The medical image is provided to a trained anatomy segmentation network. A feature map is received from the trained anatomy segmentation network. The feature map indicates the location of at least one feature within the medical image. The feature map is provided to a trained classification network. The trained classification network was pre-trained on a plurality of feature map outputs of the segmentation network. A disease detection is received from the trained classification network. The disease detection indicating the presence or absence of a predetermined disease.

Disease detection algorithms trainable with small number of positive samples

Methods and systems are directed to training an artificial intelligence engine. One system includes an electronic processor configured obtain a set of reports corresponding to a set of medical images, determine a label for a finding of interest, and identify one or more ambiguous reports in the set of repots. Ambiguous reports do not include a positive label or a negative label for the finding of interest. The electronic processor is also configured to generate an annotation for each of the one or more ambiguous reports in the set of reports, and train the artificial intelligence engine using a training set including the annotation for each of the one or more ambiguous reports and non-ambiguous reports in the set of reports. A result of the training is generation of a classification model for the label for the finding of interest.

Rule out accuracy for detecting findings of interest in images

A system for automated detection and type classification of central venous catheters. The system includes an electronic processor that is configured to, based on an image, generate a segmentation of a potential central venous catheter using a segmentation method and extract, from the segmentation, one or more image features associated with the potential central venous catheter. The electronic processor is also configured to, based on the one or more image features, determine, using a first classifier, whether the image includes a central venous catheters and determine, using a second classifier, a type of central venous catheter included in the image.

Automated detection and type classification of central venous catheters

Mechanisms are provided to implement an automated medical imaging report generator which receives an input medical image and inputs the input medical image into a machine learning (ML) computer model trained to predict finding labels based on patterns of image features extracted from the medical image. The ML computer model generates a prediction of a finding label applicable to the input medical image in terms of a finding label prediction output vector. Based on the finding label prediction output vector, a lookup operation is performed, in a medical report database of previously processed medical imaging report data structures, to find a matching medical imaging report data structure corresponding to the finding label. An output medical imaging report is generated for the input medical image based on natural language content of the matching medical imaging report data structure.

Automatic generation of medical imaging reports based on fine grained finding labels

Systems and methods generate a segmentation network for image segmentation using global optimization. A method for automatic generation of at least one segmentation network includes providing an initial set of hyperparameters to construct a segmentation network. The hyperparameters define operations for a set of block structures and connections between the block structures. The segmentation network is trained using a first set of images with ground truth. An objective function value for the trained segmentation network is generated using a second set of images having ground truth. The set of hyperparameters is updated by performing a derivative-free optimization algorithm on the objective function value to construct an updated segmentation network. The training of the segmentation network, the generating of the objective function, and the updating of the set of hyperparameters for the updated segmentation network are iterated to generate a network architecture for the segmentation network.

Network architecture search with global optimization

Systems and methods generate a segmentation network for image segmentation using global optimization. A method for automatic generation of at least one segmentation network includes providing an initial set of hyperparameters to construct a segmentation network. The hyperparameters define operations for a set of block structures and connections between the block structures. The segmentation network is trained using a first set of images with ground truth. An objective function value for the trained segmentation network is generated using a second set of images having ground truth. Generating the objective function includes setting the objective function to a predetermined value responsive to identifying an untrainable condition of the trained initial segmentation network. The set of hyperparameters is updated by performing an optimization algorithm on the objective function value to construct an updated segmentation network. The training of the segmentation network, the generating of the objective function, and the updating of the set of hyperparameters for the updated segmentation network are iterated to generate a network architecture for the segmentation network.

Handling untrainable conditions in a network architecture search

Methods for image segmentation and smoothing of image segmentations are provided. In various embodiments, a plurality of training images is provided to a segmentation network. A candidate segmentation is obtained from the segmentation network for each of the plurality of training images. Each candidate segmentation is compared to a ground truth segmentation to compute a loss metric for each candidate segmentation. Based on the gradient of the loss, the segmentation network is trained to minimize level set smoothing energy. In various embodiments, an input image is downsampled from a first resolution to a second, lower resolution. The downsampled image is provided to a segmentation network. A segmentation at the second resolution is obtained from the segmentation network. The segmentation from the second resolution is upsampled to the first resolution. The input image and the upsampled segmentation are provided at the first resolution to a convolutional network. An updated segmentation at the first resolution is obtained therefrom.

Framework for integrating deformable modeling with 3D deep neural network segmentation

3D segmentation with exponential logarithmic loss for highly unbalanced object sizes is provided. In various embodiments, an artificial neural network is trained to label an anatomical feature in medical imagery by: i) providing at least one medical image to the artificial neural network; ii) determining from the artificial neural network a predicted segmentation for the at least one medical image; iii) comparing the predicted segmentation to ground truth segmentation, and computing therefrom a loss function, the loss function having an exponential-logarithmic term; and iv) updating the artificial neural network based on the loss function.

3D segmentation with exponential logarithmic loss for highly unbalanced object sizes

Automated detection of septal defects in cardiac computed tomography images is provided. In various embodiments, a blood pool is extracted from a CTA image of a heart. The blood pool is determined to be a left pool. A septal curve of the left pool is extracted. A plurality of features of the septal curve is determined. The heart is classified as having an atrial septal defect, ventricular septal defect, or no defect based on the plurality of features.

Growing community of inventors

San Jose, CA, United States of America

Chun Lok Wong