Fuzzy logic is a concept first introduced by Zadeh in 1965 to model logical reasoning with a level of vagueness. I wanted to develop a Fuzzy Inference System (FIS) to classify breast cancer tumours.

Why use fuzzy logic and not some trendy, “state-of-the-art” approach like neural networks? Because conclusions reached through use of fuzzy logic can be easily understood and interpreted by a medical practitioner, which provides a far more “explainable A.I.” solution than black box neural nets for example.

Furthermore, the problem domain of determining any presence of a single disease may manifest itself quite differently depending on the patient, and with different intensities. A single symptom may correspond to different diseases. On the other hand, several diseases present in a patient may interact and interfere with the usual description of any of the diseases. Such overlapping between diseases seemed a perfect fit for a fuzzy logic-based classification system able to model blurriness.

Matlab’s fuzzy logic toolbox is the most popular fuzzy logic library I’m aware of, with a large user community and many documented example use cases. However, I’m a Python developer at heart, so was happy to discover the Skfuzzy Python library. Some simple implementation examples can be found here.

The Python-based FIS I developed to classify breast cancer tumours worked fantastically well. Tested with the Wisconsin Diagnostic Breast Cancer Dataset, the best FIS configurations were able to correctly predict all malignant cancer samples. It outperformed tuned Keras neural networks and other classifiers including decision trees and random forest. The code can be found here, along with the supporting paper.