AAI_2025_Capstone_Chronicles_Combined

Breast Tumor Classification Using Quantum Neural Networks

21

Increasing the features selected is another clear area to explore. Most attempts were

limited to around 10 features, since those could quickly be trained and iterated on. However,

those features could be approximately doubled- but suffered an exponential slowdown due to a

simulated quantum computer being used. Of course, a way to combat this slower iteration with

more features would be to increase the compute power used simulate the quantum computer.

The use of different quantum circuits is an area for further work. While different numbers

of hidden layers were attempted, each layer was essentially the same design as the others-

although clearly modified during training. (Refer to Figure 11 for an illustration of the quantum

circuit for a single layer.) Its possible further improvements could be made if these layers were

different, or perhaps a mix and match of different layers instead of the same.

A final area for future work would be to compare to an equivalent classical neural

network, operating on bits. While the 90% accuracy with only 10 qubits for 10 features sounds

impressive, another area to explore would be to see how that performs against the classical

equivalent. Specifically, take the same features and map them to bits, have the same number of

hidden layers and nodes, same activation functions, and so on. While it is highly likely that a

quantum neural network performs much better, due to qubits having higher dimensions than

bits, exactly how much better is unknown until this is attempted.

Conclusion

Quantum computation allows for certain problems to be solved that are not feasible to

solve on existing classical computers. Artificial intelligence is at the forefront of computer

science and classification using neural networks is a prime application. This work combines

quantum computing with artificial intelligence to perform classification of tumors that may be

cancerous or not using quantum neural networks. The results achieved exceeded expectations:

over 90% accuracy only using 10 of the 30 available features, with each feature represented by

274