94 lines (53 with data), 5.3 kB
Breast Cancer Detection
This project determines if there is breast cancer based off an ultrasound.
Algorithm
This project uses a resnet18 model that was retrained with two different sets of data. One dataset contained images of ultrasounds that were confirmed to contain malignant breast cancer. Another dataset was consisted of ultrasounds that did not show any evidence of breast cancer. Once the model was retrained, I exported it using the ONNX format. The program uses imagenet.py to determine the whether or not breast cancer is present in an ultrasound.
The dataset can be found here: https://www.kaggle.com/datasets/aryashah2k/breast-ultrasound-images-dataset
Running this project
Setup
1) Download the jetson-inference container from github to a jetson-nano: https://github.com/dusty-nv/jetson-inference
3) Change directories into jetson-inference/python/training/classification/data
4) Create a directory called "breast_cancer_detection"
5) Run this command in the terminal to download the data
`wget https://www.kaggle.com/datasets/aryashah2k/breast-ultrasound-images-dataset/download?datasetVersionNumber=1 -O ultrasounds.tar.gz`
6) Make 3 directories called "train, "test", and "val" in the "breast_cancer_detection" directory
7) Make a .txt file called "labels.txt" in the "breast_cancer_detection" directory and write down the following (in exactly that order)
breast_cancer
no_breast_cancer
8) Make 2 more directories called "breast_cancer" and "no_breast_cancer" in each directory: "train", "val", and "test"
9) Move around 5% of the images from the "normal" folder inside of the downloaded "ultrasounds" folder to the jetson-nano to the directory jetson-inference/python/training/classification/data/breast_cancer_detection/test/no_breast_cancer
10) Move around 5% of the images from the "malignant" folder inside of the downloaded "ultrasounds" folder to the jetson-nano to the directory jetson-inference/python/training/classification/data/breast_cancer_detection/test/breast_cancer
11) Move around 15%-25% of the images from the "normal" folder inside of the downloaded "ultrasounds" folder to the jetson-nano to the directory jetson-inference/python/training/classification/data/breast_cancer_detection/val/no_breast_cancer
12) Move around 15%-25% of the images from the "malignant" folder inside of the downloaded "ultrasounds" folder to the jetson-nano to the directory jetson-inference/python/training/classification/data/breast_cancer_detection/val/breast_cancer
13) Put the remaining images from the "normal" folder to the directory
jetson-inference/python/training/classification/data/breast_cancer_detection/train/no_breast_cancer
14) Put the remaining images from the "malignant" folder to the directory
jetson-inference/python/training/classification/data/breast_cancer_detection/train/breast_cancer
15) Make a new directory called "breast_cancer_detection" in jetson-inference/python/training/classification/models
Execution
1) Change directories to jetson-inference
2) Type and run ./docker/run.sh in the terminal
3) Then change directories to jetson-inference/python/training/classification
4) Now train the model by running this command in the terminal
python3 train.py --model-dir=models/breast_cancer_detection data/breast_cancer_detection --epochs=150
Note: This will take approximately 2 hours and 30 minutes
5) Once done, export the model by running this script
`python3 onnx_export.py --model-dir=models/breast_cancer_detection`
7) If you are in the docker container, exit it by pressing ctrl d
8) Change directories to jetson-inference/python/training/classification
9) In the terminal enter in
NET=models/breast_cancer_detection and
DATASET=data/breast_cancer_detection
9) Then enter in
imagenet.py --model=$NET/resnet18.onnx --input_blob=input_0 --output_blob=output_0 --labels=$DATASET/labels.txt $DATASET/test/breast_cancer/malignant8.png output.jpg
10) If you look in your classification directory there will be a file called output.jpg (or whatever you named the output to be)
11) Display the file to see if there is breast cancer
Note:
In step 4 you can replace "malignant8.png" with any file in the breast_cancer directory and you can rename "output.jpg" to anything you want.
You can also try using ultrasounds that have no breast cancer by changing "breast_cancer" in step 4 to "no_breast_cancer" and replacing
"malignant8.png" with any file in the "no_breast_cancer" directory.
If you like, you can also upload images of ultrasounds found online to the "breast_cancer" or "no_breast_cancer" directories and try running step 4
with your own image instead of "malignant8.png" to check if there is breast cancer.