01
Abstract
Breast cancer is the most commonly diagnosed form of cancer in the world and continues to claim more lives every year. Creating more effective methods of diagnosing it early is crucial, since an early diagnosis leads to a higher survival rate. Using explainable AI, clinicians can get more valuable insight into an AI model's findings. Federated learning and differential privacy combined allow these AI models to be implemented so that the risk of data leakage can be reduced. While existing literature generally examines these topics separately, this research compares the use of explainable AI methods, federated learning, differential privacy, and the associated challenges. The analysis identifies Grad-CAM as one of the most promising explainable AI methods and finds that federated learning and differential privacy are most useful when combined. The research also finds that the benefits of differential privacy can only be meaningfully determined if the relevant privacy parameters are reported transparently.
02
Research Context
The survival rate of breast cancer decreases considerably the longer it remains untreated, making it extremely important that it is found early. Tools that allow clinicians to catch breast cancer early have the potential to save many lives. Mammography, although it has some caveats, continues to generally be the best form of imaging to screen for breast cancer. Mammography is well-suited for AI use since AI models can be trained to find patterns in images.
A traditional approach to an AI detection model is to maximize accuracy, however in a clinical setting accuracy alone is not enough. Clinicians need to understand why a model made its prediction in order to verify it since these models can make mistakes. Sensitive data is also required to train these models, but research has shown that attacks can expose sensitive training data. Strong privacy measures must be put in place to reduce the potential of accidental leakage against a skilled adversary.
My research explores both of these problems together, examining explainable AI, federated learning, and differential privacy for mammogram-based cancer detection.
03
Research Questions
Three research questions guided the work:
RQ1: Which explainable artificial intelligence methods are most suitable for mammogram-based cancer detection?
RQ2: How can federated learning and differential privacy reduce privacy risks in breast cancer artificial intelligence models?
RQ3: What challenges arise in balancing accuracy, interpretability, and privacy?
I found that Grad-CAM is one of the most promising XAI methods for breast cancer screening. I also found that federated learning can be used to reduce the risk of a data leak, however combining federated learning with differential privacy can provide stronger privacy protection than federated learning alone. I also found that the privacy benefits of differential privacy can only be evaluated meaningfully if the parameters and privacy metrics are reported transparently, since many implementations of differential privacy can produce little to no meaningful privacy benefit.
04
Additional Work
Recently I've been working on a related project, training EfficientNetB0, InceptionV3, and MobileNetV2 for malignant classification with Python. I've improved the malignant-class F1 score by 13.5% over the baseline that I inherited. I'm currently introducing differential privacy to measure how different privacy budgets affect model performance.