Application of deep learning for medical imaging evaluation
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.
1. A method for streamlining tuberculosis detection by automating an X-ray screening and prioritizing process, comprising:
receiving and processing images from chest X-ray scan imaging procedure data;
detecting and localizing medical abnormalities of the images using a deep learning system, wherein the deep learning system carried out by a computer is developed by the steps of:
selecting medical imaging scans and extracting the medical abnormalities using natural language processing (NLP) algorithms, wherein the medical abnormalities comprise blunted costophrenic angle, calcification, cardiomegaly, cavity, cervical rib, consolidation, hyper inflation, fibrosis, prominence in hilar region, opacity, pleural effusion, and scoliosis;
pre-processing the selected medical imaging scans by resizing and tag-specific data augmentations;
training a deep learning algorithm with the selected medical imaging scans;
predicting the presence or absence of a particular type of medical abnormalities by combining the predictions of multiple classification models using different initial conditions, architectures and sampling via a majority ensembling scheme, wherein the multiple classification models are pre-trained on a task of separating chest X-rays from X-rays of other body parts; and
generating a score that corresponds to a level of recognition of each of the medical abnormalities and outputting a bounding box representing the precise location and extent of the medical abnormalities; and
generating another score and advice that corresponds to a level of recognition of tuberculosis based on the scores of the medical abnormalities that suggest typical or atypical tuberculosis.
2. The method of claim 1 , wherein the deep learning algorithm comprises a convolutional neural network architecture.
3. The method of claim 2 , wherein the architecture is modified by pre-training on the task of separating chest X-rays from X-rays of other body parts.
4. The method of claim 1 , wherein the models are selected using various heuristics.
5. The method of claim 1 , wherein the accuracy of the deep learning algorithm for detecting the medical abnormalities is validated by comparing with radiologist reports.
6. A system for streamlining tuberculosis detection by automating an X-ray screening and prioritizing process, comprising:
a computing device configured to receive and process images from chest X-ray scan imaging procedure data;
the computing device configured to detect and localize medical abnormalities of the images using a deep learning system, wherein the deep learning system is developed by the steps of:
selecting medical imaging scans and extracting the medical abnormalities using natural language processing (NLP) algorithms, wherein the medical abnormalities comprise blunted costophrenic angle, calcification, cardiomegaly, cavity, cervical rib, consolidation, hyper inflation, fibrosis, prominence in hilar region, opacity, pleural effusion, and scoliosis;
pre-processing the selected medical imaging scans by resizing and tag-specific data augmentations;
training a deep learning algorithm with the selected medical imaging scans;
predicting the presence or absence of a particular type of medical abnormalities by combining the predictions of multiple classification models using different initial conditions, architectures and sampling via a majority ensembling scheme, wherein the multiple classification models are pre-trained on a task of separating chest X-rays from X-rays of other body parts; and
generating a score that corresponds to a level of recognition of each of the medical abnormalities and outputting a bounding box representing the precise location and extent of the medical abnormalities; and
the computing device configured to generate another score and advice that corresponds to a level of recognition of tuberculosis based on the scores of the medical abnormalities that suggest typical or atypical tuberculosis.
7. The system of claim 6 , wherein the deep learning algorithm comprises a convolutional neural network architecture.
8. The system of claim 7 , wherein the architecture is modified by pre-training on the task of separating chest X-rays from X-rays of other body parts.
9. The system of claim 6 , wherein the models are selected using various heuristics.
10. The system of claim 6 , wherein the accuracy of the deep learning algorithm for detecting the medical abnormalities is validated by comparing with radiologist reports.
11. The system of claim 6 , wherein the algorithms achieve Area under the Receiver Operating Characteristic curve (AUC) of 0.93±0.01 for detection of abnormal scans, and AUCs of 0.94±0.02, 0.88±0.03, 0.97±0.02, 0.93±0.07, 0.93±0.04, 0.88±0.05, 0.89±0.05, 0.93±0.02, 0.98±0.02 for the detection of blunted costophrenic angle, calcification, cardiomegaly, cavity, consolidation, fibrosis, hilar enlargement, opacity and pleural effusion, respectively.