Based on the study of lung infection detection technology, can detect the prevalence of medical imaging -X-rays through artificial intelligence technology to effectively improve the diagnosis accuracy and efficiency; this contest problems mainly on the chest X-ray results sick classification of 4 diseased categories, data set into a ratio of 6: 2: 2;
problem analysis
First View original picture visual features to note here on CNN and the intuitive nature of man there are differences, it is not necessarily effective for data enhancement on intuition.
- The picture does not appear flipped vertically
- Strictly speaking there will be flipped horizontally
- Picture is scaled to 224, 224, you may have lost some of the information
- Belong to a single-channel grayscale images
Analysis of data labels, and number:
- Data are divided into four categories
- The amount of data about the training of about 18,000
- There is imbalance data
Data enhancement
By observing the data, the following data enhancement means
- Flip Horizontal, strictly not allowed, but considering the incomplete symmetry around the lungs
- A slight rotation
- Small amplitude panning
- Small amplitude scaling
Model Design
Initial depth model uses fewer Resnet34, fast training speed, has been employed in subsequent experiments, in order to suppress over-fitting, before the final classification added Dropout = 0.2, slightly improved performance.
This problem model to enhance the accuracy of the more obvious tips
- TTA prediction process used, test time is enhanced, and the results averaged after weighting enhanced direct
- Fusion model, a model that does not start up the fusion point obvious, but the key point up after use discovery.
Experience Sharing
Began last September, template platform that allows fast learning on the phone, the main purpose of the contest is to enhance their own, this process is the ability to exercise their own code and the ability to check documents.
See previous share, from the beginning to find papers related fields, and then step by step to enhance the accuracy ..
This article is the author FlyAI platform release of original content, using the Creative Commons Attribution - Noncommercial - No Derivative Works 4.0 International License Agreement for licensing, reproduced, please attach the original source link and this statement.
This link address (direct video to explain): https://www.flyai.com/n/116619
