1 of 10

IDC IDENTIFIER

Anthony Martin

2 of 10

USE CASE

  • The primary aim of the project is to use machine learning to assist in the identification of Invasive Ductal Carcinoma (IDC) from images taken as extracts from whole mount slides.
  • The goal of the project is to assist analysts and researchers in rapid identification of the exact location of IDC cells.
  • Successful deployment of the process is defined by a reduction in processing time for patient results, saving time for both researchers and patients.

3 of 10

WHY DEEP LEARNING?

  • Allows for a machine to begin learning through training sets and make decisions without extensive feature implementation.
  • Works best for unstructured data, such as pictures, sound, and language.
  • More accurate when more data is used.

4 of 10

DATA SET ACQUISITION AND ASSESSMENT

  • Initial data set is available publicly through a repository on Kaggle.com.
  • Full citations can be found in the appendix.
  • As the dataset is maintained and hosted by Kaggle, the was little additional preprocessing work that was required. Though the notebook also must be run through Kaggle, or both it and the dataset will need to be downloaded and run locally.

5 of 10

DATA EXPLORATION

  • Ensuring that each image had a corresponding label, and was a correct image was the first step in exploring the dataset.

​

  • Then the dataset was checked for size and categorical anomalies, in both the training and validation sets.

6 of 10

ACCURACY BEFORE AND AFTER NORMALIZATION

  • Feature normalization stabilizes and enhances model accuracy, and as datasets grow larger, allows the model to run more efficiently.
  • By the 15th epoch, model accuracy was the same for both the normalized and unnormalized versions.
  • The normalized version was 10ms faster per step and will perform faster with larger datasets.

7 of 10

CONFUSION MATRIX

  • A confusion matrix shows the differences between the predicted results and the true results, allowing for validation of the model’s ability to perform.
  • In our test sample, only 3% were misidentified.

8 of 10

DEEP LEARNING VS LOGISTIC REGRESSION

The two confusion matrices for the Deep Learning model (left) and the Logistic Regression model (right) are used to compare the performance of the two models.

9 of 10

CONCLUSIONS

The accuracy of the model allows for much faster identification for researchers, as well as the ability to handle much larger sets of information.

​

The pinpointing of the invasive cells will expedite assigning a level of aggression to the disease and grant much needed time.

10 of 10

APPENDIX