Chest X-Ray Analysis Using Deep Convolutional Network
Principal Investigator
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCO-302
Initial CDAS Request Approval
Sep 8, 2017
Title
Chest X-Ray Analysis Using Deep Convolutional Network
Summary
Chest X-Rays are widely used for diagnosing abnormalities in heart and lung area. Automatic detection of these abnormalities from the X-Rays would greatly enhance real world diagnosis process and would guide the radiologists to make more reliable diagnosis. In our recent paper [1], we investigated publicly available Indiana chest X-Ray dataset in order to detect and localize abnormalities in the heart and lung area. The number of samples in the dataset is low. Thus we have proposed methods that could classify accurately in such small datasets and have reported a staggering 17% improvement in accuracy in Cardiomegaly detection over the existing rule-based methods. We have explored various neural network configurations and identified the network/pipeline that achieves the highest accuracy for chest X-Ray abnormality detection. For the first time in literature, we have reported the detection accuracy, AUC, sensitivity and specificity of 20 different abnormalities in chest X-Ray. This work will form the benchmark with which future studies can be compared. Besides detection, we have shown that our models are capable of localizing the abnormalities in the chest X-Ray when the abnormalities are spatially spread out. In addition, deep learning based models learn features that are often different than radiologist’s intuition, hence provides a new way of thinking about the abnormalities.
With the success of our previous study on Indiana dataset, we want to evaluate the performance of our best machine learning pipelines on larger and more representative datasets. We want to test at least three hypotheses utilizing the PLCO dataset: i) Does the accuracy of the machine learning models improve when fed with more data? ii) We have setup a new detection accuracy benchmark for the Indiana dataset. Can we do the same for PLCO dataset? Iii) Combined with the clinical information in PLCO dataset, how well does the models perform in retrospective studies of clinical decision-making. Also, the PLCO dataset includes some abnormalities that are not present in the Indiana Dataset. We will setup accuracy benchmark studies for these abnormalities as well. The PLCO dataset can thus help us answer our hypotheses questions and move the field of automatic chest X-Ray reading research forward with larger and more representative studies.
[1] Mohammad Tariqul Islam, Md. Abdul Aowal, Ahmed Tahseen Minhaz, Khalid Ashraf, "Abnormality Detection and Localization in Chest X-Rays using Deep Convolutional Neural Networks", aRxiv pre-print: arXiv:1705.09850, 2017.
With the success of our previous study on Indiana dataset, we want to evaluate the performance of our best machine learning pipelines on larger and more representative datasets. We want to test at least three hypotheses utilizing the PLCO dataset: i) Does the accuracy of the machine learning models improve when fed with more data? ii) We have setup a new detection accuracy benchmark for the Indiana dataset. Can we do the same for PLCO dataset? Iii) Combined with the clinical information in PLCO dataset, how well does the models perform in retrospective studies of clinical decision-making. Also, the PLCO dataset includes some abnormalities that are not present in the Indiana Dataset. We will setup accuracy benchmark studies for these abnormalities as well. The PLCO dataset can thus help us answer our hypotheses questions and move the field of automatic chest X-Ray reading research forward with larger and more representative studies.
[1] Mohammad Tariqul Islam, Md. Abdul Aowal, Ahmed Tahseen Minhaz, Khalid Ashraf, "Abnormality Detection and Localization in Chest X-Rays using Deep Convolutional Neural Networks", aRxiv pre-print: arXiv:1705.09850, 2017.
Aims
The specific aims of the project are:
1. To use Deep Convolutional Neural Networks to find and localize abnormalities in Chest X-Rays.
2. Investigate if/how more training data can improve machine learning detection accuracy.
3. Setup abnormality detection benchmarks on the PLCO dataset using our high accuracy machine learning models reported in [1].
4. If gold standard ground truth labels are present through clinical information/biopsy/culture, then we will evaluate the performance of clinical decision-making for various abnormalities.
Collaborators
Khalid Ashraf, Semion Inc
Mohammad Tariqul Islam, Bangladesh University of Engineering and Technology & Semion Ltd
Mohammad Abdul Aowal, Semion Ltd
Ahmed Tahseen Minhaz, Semion Ltd
Tanveerul Islam, Semion Ltd