AAI_2025_Capstone_Chronicles_Combined

Model

Accuracy

F1 Score

Precision

Recall

Custom CNN w/ Real Images

.71

.65

.80

.71

Custom CNN w/ Real Images -Balanced

.76

.72

.81

.76

VGG w/ Real Images

.78

.72

.87

.71

VGG w/ Real Images – Balanced

.82

.78

.88

.76

Custom CNN w/ ACGan Images

.80

.80

.82

.80

VGG w/ ACGAN Images

.81

.81

.82

.84

Custom CNN w/ StyleGAN

.86

.86

.86

.86

VGG w/ StyleGAN images

.92

.91

.92

.92

Table 1. CNN model performance metrics, sorted by F1 score

The hypothesis that synthetic data augmentation would enhance model

performance is strongly supported by these findings, particularly in the marked

improvements demonstrated by GAN-based augmentation techniques. The increased

accuracy and balanced precision and recall observed in GAN-augmented models,

especially the VGG model trained on StyleGAN-generated images, suggest significant

potential to enhance diagnostic capabilities, reduce misdiagnoses, and assist healthcare

professionals in clinical settings. Additionally, eliminating the need for further patient data

mitigates concerns around patient privacy, data security, and the overhead associated with

data collection.

117