#Hip implant classifier
A common problem among orthopedic surgeons is the recognition of patients’ hip implants through an x-ray. Given the number of implants available in the market, recognizing them becomes a difficult task, especially for traumatologists with few years of experience. That’s why, in conjunction with my brother -who is a doctor-, I started to build an application to recognize hip implants.
To achieve a good recognition, I decided to solve 3 tasks:
- Hip implant detector in the image
- Align the implant
- Apply a classifier to the final images They were solved using CNN’s for each task.
Hip implant detector
Detecting the implant in the image was the first task to be solved. For this, the “tensorflow” library was used, and a detector was trained with manually tagged images.
The results of the detector were highly satisfactory
Hip implant aligner
After obtaining the bounding box of the image, a CNN was trained to find the ends of the implant and thus align the image.
To align the image, the following architecture was used.
Hip classifier
After obtaining a straightened image, a classifier with Google’s inception architecture was used.
Future work
I am currently testing other architectures and testing filters to improve the accuracy of the model, because CNN’s currently identify and learn from other parts of the X-ray that are not of interest (no hip implants).