Predicting Breast Cancer Events in Ductal Carcinoma In Situ (DCIS) Using Generative Adversarial Network Augmented...

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Standard clinicopathological parameters offer limited predictive value for recurrence in ductal carcinoma in situ (DCIS).
• Accurate prediction of breast cancer events (BCEs) is crucial for tailoring treatment strategies, avoiding overtreatment in low-risk patients and guiding aggressive therapy for high-risk individuals.
• Generative adversarial networks (GANs) are advanced deep learning models used for generating high-quality synthetic images.

Purpose of the Study:

• To develop and validate a deep learning (DL) classification network for predicting BCEs in DCIS patients.
• To assess the efficacy of DL models trained on both real and GAN-generated H&E images for recurrence prediction.
• To evaluate the independence of the DL model's predictions from traditional clinicopathological factors.

Main Methods:

• A DL classification network was developed using hematoxylin and eosin (H&E) stained image patches from DCIS cores.
• The model was trained on a dataset including both actual patient data and GAN-generated image patches.
• Model performance was evaluated on a hold-out validation set (n=66), with AUC calculated. Bayesian analysis was employed to confirm independence from clinicopathological parameters.

Main Results:

• The DL classification model achieved an Area Under the Curve (AUC) of 0.82 on the hold-out validation dataset.
• Bayesian analysis demonstrated that the DL model's predictions were independent of standard clinicopathological parameters.
• The study successfully utilized GAN-generated images to augment the training dataset for the DL model.

Conclusions:

• Deep learning models analyzing H&E images show promise as a risk stratification tool for DCIS patients.
• This AI-driven approach can potentially personalize therapeutic decisions, distinguishing between high-risk and low-risk patients.
• DL models offer a novel strategy to improve the accuracy of predicting breast cancer events in DCIS, complementing existing methods.