|
---
|
|
license: mit
|
|
language:
|
|
- en
|
|
tags:
|
|
- image classification
|
|
- classification
|
|
- medical imaging
|
|
- medical
|
|
- dicom
|
|
- cancer
|
|
metrics:
|
|
- '62% Sensitivity'
|
|
---
|
|
|
|
# HerBreastsFriend(HBF)
|
|
|
|
![Demo](./assets/imgs-preview.gif)
|
|
|
|
A model for identifying breast cancer in patients inspired by a study conducted by Duke & blogged about by jamanetwork[^1].
|
|
|
|
Their studies finding's were that there's a lot of room for improvement. They came to this conclusion after building their
|
|
own AI model for breast cancer detection/prognoses and achieved a 65% on sensitivity.
|
|
|
|
### Details
|
|
|
|
![Demo](./assets/matrix-previews.gif)
|
|
|
|
- KNN strategy
|
|
- n_neighbors=5
|
|
- StandardScaler
|
|
- PCA
|
|
- n_components=2
|
|
- Trained on limited dataset(1997 images)
|
|
- I had to limit the number of data points in my model because my machine kept freezing. WIP on a solution.
|
|
- Hosted by the amazing cancerimagingarchive[^2]
|
|
|
|
### Classification Report
|
|
|
|
The initial release of HBF scored the following in our classification. 62% for average weighted across all features. A lot of room for improvement.
|
|
|
|
```sh
|
|
precision recall f1-score support
|
|
|
|
Normal 0 0.62 0.80 0.70 956
|
|
Actionable 1 0.61 0.58 0.59 760
|
|
Benign 2 0.69 0.07 0.12 164
|
|
Cancer 3 0.47 0.08 0.13 117
|
|
|
|
accuracy 0.62 1997
|
|
macro avg 0.60 0.38 0.39 1997
|
|
weighted avg 0.61 0.62 0.58 1997
|
|
```
|
|
|
|
### FAQ
|
|
|
|
I'm considering making this open source. If you'd like to contribute please give a star to let me know there's others interested.
|
|
|
|
[^1] Duke Study https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2783046
|
|
[^2] [cancerimagingarchive https://www.breastcancer.org/facts-statistics](https://www.cancerimagingarchive.net/collection/breast-cancer-screening-dbt)
|
|
|
|
|
|
|
|
|
|
A study conducted by Duke University Health System assessed "deep learning" and "medical imaging in general" have significant advancements left to go in the future.
|
|
|
|
|
|
|
|
Their conclusion comes following thier own AI models, trained to detect cancer in a non-invasive way(requiring no biopsy), was evaluated at only 65% sensitivity.
|
|
|
|
|
|
|
|
Although in reality no easy feat, a disappointing statistic from the US's 7th best University.
|
|
|
|
|
|
|
|
I, having experience in the industry & seeking a meaningful project to work on, felt compelled to see what I could do to move the needle.
|
|
|
|
|
|
|
|
The result of this was creating a model which was evaluated at 62% using scikit-learn's classification report.
|
|
|
|
|
|
|
|
The model is hosted on Hugging Face linked below. And soon radiologist & patients will all be able to use this model for free(and future improved versions of it) at https://lnkd.in/eTWCD2wu
|
|
|
|
|
|
|
|
https://lnkd.in/eTx_sw9k |