Method and system for medical imaging evaluation

Inventors

Putha, Preetham • Tadepalli, Manoj • Reddy, Bhargava • Raj, Tarun • Jagirdar, Ammar • Rao, Pooja • Warier, Prashant

Assignees

Qure AI Technologies Pvt Ltd

Abstract

This disclosure generally pertains to methods and systems for processing electronic data obtained from imaging or other diagnostic and evaluative medical procedures. Certain embodiments relate to methods for the development of deep learning algorithms that perform machine recognition of specific features and conditions in imaging and other medical data. Another embodiment provides systems configured to detect and localize medical abnormalities on medical imaging scans by a deep learning algorithm.

Core Innovation

The invention provides a method and corresponding system for developing and automating detection and localization of medical abnormalities on chest X-ray imaging scans using a deep learning algorithm. Chest X-ray imaging scans and corresponding radiology reports are selected, and medical abnormalities are extracted via Natural language processing (NLP) algorithms to generate extracted findings. The extracted findings are used as labels for training the deep learning algorithm to produce abnormality detection outputs.

The approach segments each selected chest X-ray imaging scan using an anatomy segmenter to generate segmentation masks corresponding to the chest cavity, lungs, diaphragm, mediastinum and ribs. A region of interest (ROI) generator outputs a plurality of ROIs relevant for detecting a particular abnormality, where the ROI generator uses the chest X-ray imaging scan at full resolution together with the anatomy segmentation masks. This links thoracic anatomical structure to abnormality-relevant regions for downstream detection.

Abnormalities are detected via an abnormality detector that is a hybrid classification plus segmentation network. For each ROI, the abnormality detector outputs a low-resolution probability map and a confidence score per each ROI, and it also outputs a confidence score for an entire chest X-ray scan by combining the confidence scores per each ROI. The scan-level confidence is produced by combining ROI confidence scores, as described in the patent content.

Claims Coverage

The independent claims are directed to a method for developing a deep learning system and a system that automates detection and localization. Across the independent claims, the inventive features center on NLP-extracted radiology-report findings used as training labels, anatomy segmentation that defines ROI generation, and a hybrid classification-plus-segmentation abnormality detector that outputs per-ROI probability maps and confidence scores aggregated into a scan-level confidence.

Taken together, the independent claims require an end-to-end pipeline where NLP-extracted radiology findings provide training labels, an anatomy segmenter generates segmentation masks for thoracic structures, an ROI generator uses those masks with full-resolution scans to define abnormality-relevant regions, and a hybrid classification-plus-segmentation abnormality detector produces per-ROI probability maps and ROI confidence scores that are combined into a scan-level confidence. Dependent claim coverage further specifies convex LogSumExp (LSE) convex pooling for confidence aggregation and weighted cross-entropy across three levels.

Stated Advantages

Documented Applications