This project presents a work plan and information from literature records on the development of a system to support the detection of identifiable medical conditions from chest radiographs. The detection process involved performing spatial localization and classification of medical conditions on medical images (chest radiographs), even in cases where more than one condition was found per image, or the same condition in different regions of the radiograph.
For the development of the system, the use of computer vision techniques based on deep artificial neural networks was proposed. The development of this type of system has become more relevant in recent years. The growing demand for radiologists with specialized skills, as well as the emergence of conditions related to respiratory diseases have driven the need for research and development of this type of tools.
The objective of this project has been to develop a tool to support the specialist in medical diagnosis and treatment. However, it is discarded that the information generated by the tool is considered as a final diagnosis.
Project Goals
To design and implement a system for the classification and localization of conditions with visual evidence in chest radiographs with postero anterior (PA) view, using deep learning techniques, in order to serve as a support system for specialists in the diagnosis of various types of conditions. the diagnosis of various types of conditions.
- Conduct an exploratory and quantitative analysis of publicly available chest radiograph datasets. Perform data cleaning to use only data with the required specifications.
- Apply synthetic data augmentation techniques that do not affect model interpretation.
- Train models for classification and localization of conditions considering only chest X-ray data sets.
- Train models for classification and localization of conditions considering a transfer of learning scheme.
- Perform a comparative analysis evaluating the performance of the proposed methods.
Project Requirements & Specifications
The project has been developed using the programming language Python and the interface for machine learning scheemes PyTorch. All requirements and code is available in this Github Repository.
