AAI_2025_Capstone_Chronicles_Combined

Breast Tumor Classification Using Quantum Neural Networks

13

the underlying probability distributions of the input (Karthinkeyan, Akila, Sumathi, & Poongodi,

2025).

In addition to the MNIST handwritten example this project was motivated by, there are

additional applications utilizing quantum neural networks. Vehicle classification is another

example. (Yu & Ma, 2008) have done simulations that have shown that quantum neural

networks significantly reduced training times, while still providing enough accuracy. Quantum

convolutional neural networks are a related area of active research and development

(TensorFlow, 2025).

Experimental Methods

Quantum neural networks were utilized for this project, with some components largely

remaining static and others being adjusted to see how they impacted model performance. The

largely static components included the following:

 The loss function, which was based on hinge accuracy (Pattanayak, 2021). This

included mapping the targets from 0 and 1 to -1 and 1

 At each layer, a single node for each feature plus a readout. This results in n + 1 qubits

at each layer for n features

 Use of the Adam optimizer included in TensorFlow, with default parameters (TensorFlow,

2024)

 A standard data set split of 80% training and 20% testing

Other components were adjusted with the aim of improving model performance. This

typically involved changing one to what seemed to be peak performance for that variable, then

adjusting others. An obvious area to expand on would be to take a more methodical approach

over a wider range: trying all combinations. Different parts of the model that were adjusted

included:

266