Hardware components | ||||||
| × | 1 | ||||
| × | 1 | ||||
Software apps and online services | ||||||
 |
| |||||
| ||||||
| ||||||
Hand tools and fabrication machines | ||||||
 |
| |||||
 |
| |||||
The idea is to use computer vision to recognize a tumor using FOMO, but since the images are taken by special MRT medical equipment, a camera is not needed and the algorithm can work even on devices that do not have a camera interface. This may be a small USB device, which, when receiving a photo, will return the already recognized photo back.
Data collectionAll photos of the collection are taken from the Internet, as there is no other source for obtaining such images.
I wanted to try on a controller with little RAM and got 84% accuracy
I started testing but it did not suit me, it was decided to improve the model
RetrainingThe result became a little better, but since accuracy is important here, I decided to try it on stm32f411ce for the sake of interest and then get a slightly larger but more accurate model
I added more data to get a good result
In the end I got an accuracy of 92.9%, which is not bad, considering that the training lasted a little less than 20 minutes, with a longer training the result will be higher than the image size of 64x64 grayscale, since the image was taken on special medical equipment.
When testing, the accuracy shows below, since I added some photos from a PC that did not see my model.
Images added by me without labels
The model was tested on STM32 (slightly inaccurate, but requires less memory) and on ESP32 CAM (more accurate model), and can also be used on other 128+ RAM boards both on boards that have a camera, and on microcontrollers that do not have an interface to connect a camera (pictures are taken by special medical equipment) for example STM32 F4, L4 series on ESP32 without a camera, Raspberry pi pico and then. As a result, it becomes possible to have a small device that can be connected to a PC, mobile device, or create a web server that will receive an image model, recognize and send back a ready-made image, and this will help doctors where there are no doctors of this specialty.
Edge impulse project
Comments
Please log in or sign up to comment.