AI for Disease Detection from Medical Images

A dynamic self-learning medical image network system that uses deep learning algorithms to analyze medical images and learn to interpret them without manual programming. The system involves multiple nodes, like medical imaging devices, that send user interactions with initial images to a central brain. The brain trains a deep learning algorithm on the interactions and sends it back to the nodes. Nodes apply the algorithm to subsequent images to generate results. The brain receives subsequent interactions and modifies the algorithm if they meet a confidence threshold. This allows the system to learn and improve from user feedback without manual intervention.

Source

Automatically segmenting anatomical structures in medical images with sharp cutoffs using two neural networks. The first network finds accurate cutoff planes or regions around structures. The second network uses cropped images from the first network to segment the structures with sharp edges. This avoids the rounding issues in segmentation when structures have incomplete data at the edges. The two-step process trains separate networks to better approximate the cutoff planes and structure contours from complete images.

Source

Using AI to detect medical conditions from lower resolution images like X-rays by training a model on annotated X-rays with corresponding higher resolution images like CT scans. The AI learns to identify features indicative of conditions in lower resolution images without needing higher resolution images. This enables diagnosing from X-rays without contrast or higher resolution scans. The model is trained by annotating X-rays based on features seen in corresponding CT scans.

Source

Endoscope processor that improves lesion detection accuracy by generating multiple disease statuses using separate image processing paths. The processor acquires an endoscope image, generates a first processed image using one image processing path, generates a second processed image using another image processing path, and outputs disease statuses using a learning model based on both processed images. This allows detecting lesions from both image processing paths and potentially improving detection accuracy compared to using just one processed image.

Source

Computer-aided evaluation of medical images like CT scans to diagnose diseases like coronary artery disease. The method involves segmenting and labeling the anatomy using algorithms, then applying rules to derive diagnosis information from the segmentation and feature extraction. An interactive presentation of the results allows users to see and interact with the segmentation and feature data. This provides understandable, explainable evaluation results compared to black box AI outputs. The interactive presentation allows users to modify and understand the segmentation and feature extraction that led to the diagnosis.

Source

Diagnosis assistance system that provides objective diagnosis information using multiple machine learning models. The system acquires multiple learned models from machine learning using image and diagnosis data. It calculates diagnosis information by applying these models to a new image. This provides more objective diagnosis information compared to using a single model, as the calculated diagnosis is based on multiple perspectives.

Source

Accurately detecting areas of the body in medical images indicating the presence of pneumonia. The method involves optimizing a pneumonia area detection model using category weighting coefficients and a Focal Loss function. This improves the accuracy of identifying the specific regions in medical images that show signs of pneumonia. By optimizing the detection model, it reduces the workload of later identifying pneumonia from the initial detection step.

Source

A system for analyzing pathologies using quantitative imaging to improve disease diagnosis and management. The system uses deep learning techniques on medical images to extract quantitative features that provide objective measurements of tissue characteristics. This allows more accurate and personalized diagnosis and prognosis compared to subjective visual assessment. The system can also quantify functional parameters like blood flow and perfusion non-invasively. By leveraging the power of quantitative imaging and AI, it aims to improve medical decision-making, reduce costs, and prevent overtreatment by providing objective measurements of disease progression and response to therapy.

Source

Real-time abnormality detection and notification during medical imaging scans to help diagnose patients faster and more accurately. The system uses AI to analyze images in real-time as they are acquired and detect potential abnormalities. It then presents the user with a notification containing details of the potential abnormality and options to accept, postpone, or reject the notification. If accepted, a report is immediately generated. If postponed, the report is delayed until after the scan. If rejected, the abnormality is logged but not reported. This allows the user to focus on scanning while potentially important findings are flagged without interrupting the flow.

Source

A system for accurately screening for eye diseases like glaucoma by linking time-stamped OCT scan data from multiple scans to a patient's biometric identifier. This allows comparing OCT scans from different time points to detect changes indicative of eye diseases like glaucoma, even if the patient changes clinics or machines. The system involves distributed OCT scanners capturing OCT and biometric data, transmitting it to a central database that links the scans to the biometric data for a patient. AI can then analyze the linked scans to diagnose eye diseases.

Source

Discriminating properties of structures in medical images without needing large labeled datasets. The method involves using a trained neural network to initially score properties of a structure in an image. It then corrects the property scores for some items based on a relationship matrix derived from analyzing co-occurrence of properties in medical reports. This matrix has weights that reflect how often certain property pairs occur together. The corrected scores are used to make the final property discrimination. The neural network is still trained using supervised data, but the relationship matrix is derived separately. This allows leveraging co-occurrence patterns in reports to enhance property discrimination without requiring labeled examples for every property.

Source

Method for processing breast tomosynthesis images to improve accuracy of lesion detection in 3D mammography. The method involves using a two-stage convolutional neural network (CNN) followed by a recurrent neural network (RNN) to process 3D breast images. The CNN extracts features from the images and generates a lesion mask. The RNN combines cropped heat maps from the CNN at different image levels to generate an integrated heat map. This integrated heat map is then used to train the RNN. The two-stage CNN reduces errors, and the RNN extracts relationships between image levels.

Source

Automated measurement tool for evaluating liver diseases like non-alcoholic fatty liver disease (NAFLD) using ultrasound images. The tool determines if an ultrasound frame has the correct view and sufficient quality, then measures visceral fat thickness. It uses AI models trained on images with progressive quality reductions to learn adequate image quality thresholds. This reduces operator variability by standardizing image requirements. The tool also assesses liver disease risk based on visceral fat measurements.

Source

Automated detection of ophthalmic diseases from images of the eye using a weighted ensemble of hierarchical end-to-end models trained on ophthalmic images. The models are trained on a labeled dataset and ranked based on their performance. Weights are assigned proportional to rank. When presented with a new image, the class prediction of each model is computed. The weighted average of class scores is taken as the ensemble class score. This weighted averaging uses rank-based weights to improve generalization compared to blindly averaging predictions.

Source

Using AI to extract and fuse mammographic features from multi-view mammograms to improve breast cancer screening. The technique involves training a neural network to extract features from single-view mammograms, then using attention mechanisms to combine and fuse the features from multiple views into a representation that can better detect breast abnormalities. The neural network architecture includes self-attention modules for single-view feature extraction, cross-attention modules for combining features from different views, and decoding modules to generate final predictions. The network is trained on mammograms with known abnormalities to learn how to extract and fuse the features that indicate cancer.

Source

AI-assisted medical imaging and report generation system that integrates AI algorithms into a unified user interface to improve efficiency and accuracy of medical image interpretation and reporting. The system provides AI-assisted image segmentation and labeling, AI-assisted dictation of findings, bi-directional dynamic linking of findings, AI finding display and interaction, and AI-enabled quality metrics. It generates reports with AI-assisted findings that can be accepted by the user and automatically incorporated into the report. The system aims to provide a complete radiology workflow with integrated AI functionality.

Source

Automated detection of anatomical landmarks from 3D medical images using deep learning, without user input or separate 3D model extraction. The method involves projecting the 3D image into 2D intensity projections, detecting an initial landmark using a 2D convolutional network, then using a different 3D convolutional network on a volume area centered on the initial landmark to find the detailed landmark. This allows accurate and automated landmark detection from 3D medical images with low quality and metal artifacts, as well as landmarks not on bone surfaces.

Source

Using machine learning and AI to analyze tissue samples more efficiently and accurately for medical diagnosis and surgical planning. The system receives whole slide images of tissue sections, determines if each section contains complete tissue versus missing areas, and identifies if sections have tumor versus healthy tissue. It generates a reconstructed model of the removed tissue with mapped tissue types. This provides a visual representation of the tissue sections and components that can be used for diagnosis, surgery planning, and research analysis. The machine learning is trained using labeled tissue images to learn to classify sections.

Source

Automated histopathological image analysis using machine learning to aid medical diagnosis. The method involves training a machine learning algorithm on histopathological images by analyzing individual tiles instead of whole images. This allows using large datasets of images with varying outcomes. The algorithm learns to predict outcomes for new images by aggregating scores from tiles. Thresholding scores provides a single outcome. This enables training and analyzing images from different scanners. The aggregated tile scores provide useful analysis even if individual tiles don't directly indicate outcomes.

Source

Automatically determining medical recommendations for patients based on multiple medical images from different modalities using a machine learning classifier. The system maps regions of interest from images of different modalities related to a patient's condition, generates annotations and clinical data for each ROI, inputs it into the classifier, and uses it to determine a medical recommendation for the patient's condition. This allows integrating and analyzing images from different modalities to provide more informed and comprehensive medical recommendations.

Source