AAI_2025_Capstone_Chronicles_Combined

Breast Tumor Classification Using Quantum Neural Networks

16

Results and Conclusion

There were several steps that had to be performed to take the existing classical data set

and transform it for use in a quantum neural network. As previously mentioned, the simulated

quantum computer utilized in Cirq has a limitation of ~20 qubits. This meant that feature

selection could include at most about 20 features. Each data point has three variations (mean,

worst, standard error), which were each tried as individual representations of the data. Another

area explored with regards to feature selection was trying to pick different features from the cold

spots in the correlation matrix. For example, attempting the area mean combined with the worst

fractal dimension (correlation of 0.0071). As previously mentioned, simulated quantum computer

experiences and exponential slow down as qubits are added. This slow down was most

apparent in this step. Specifically, selecting 10 features resulted in a training time of 23 seconds

in one case while doubling that to 20 features took about 65 minutes in another case.

Quantum feature mapping is a step unique to quantum neural networks. For this, each

normalized feature was mapped to the quantum state of |0> or |1> , essentially the quantum

equivalent of 0 and 1. This represents a significant down sampling of the data, on top of the

reduction in features already mentioned. As one would suspect, this had a tremendous negative

impact on model performance, with some trials performing no better than guessing the

classification of cancer or not- and in some cases worse.

269