Breast Cancer Detection Model
Description
The Breast Cancer Detection Model (BCDM) is a so-called deep learning model, which has been trained on a mammogram dataset made available by the Radiological Society of North America (RSNA). The mammograms were provided in the form of DICOM-files. The dataset consisted of some 54,000 mammograms, 2.12% of which were confirmed breast cancer cases.
To ensure on un-biased training of the model, the mammograms classified as malignant were duplicated using random augmentation. The final preprocessed dataset consisted of some 100,000 mammograms, 70% of which were used for training, 15% for validation and the remainder for testing.
The test results are shown in the below table.
| Metric | Value |
|---|---|
| Accuracy | 0.956530 |
| Error Rate | 0.043469 |
| Precision | 0.971376 |
| Recall (Sensitivity) | 0.939685 |
| False Positive Rate | 0.027027 |
| True Negative Rate (Specificity) | 0.972973 |
| F1-Score | 0.955268 |
| AUC | 0.976541 |
| Log Loss | 0.174501 |
Although the test result shown above are excellent for industry standards, the model has not been medically validated nor approved by any health care authority. The model is for demonstration purposes only and by no means intended to be medical advice nor diagnosis.
Mammogram
Usage
To use the model, select a mammogram as image (png, jpg or tiff) using the Select Image button. As the model runs in the browser, the selected image will not be uploaded to any server. This ensures privacy and avoids latency. After the image is selected, the mammogram will be shown in the right window and the Clear Image and Predict buttons are clickable. Within a few seconds the model returns the prediction result. Please note the processing time depends on your internet speed.