Generalization error analysis for deep convolutional neural network with transfer learning in breast cancer diagnosis

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Healthcare
• Deep Learning in Radiology

Background:

• Deep convolutional neural networks (DCNNs), a form of artificial intelligence (AI), offer advancements over traditional computer-assisted diagnostic tools in medical imaging.
• Training DCNNs typically requires large datasets, which are often limited in medical imaging tasks, necessitating the use of transfer learning.
• Automatic data mining can expand datasets but may introduce noise, such as incorrect labels or misclassified images.

Purpose of the Study:

• To investigate the generalization error of DCNNs with transfer learning in medical imaging, specifically for classifying malignant and benign breast masses on mammograms.
• To analyze the impact of noisy labels and corrupted data within the training set on DCNN performance.
• To understand the trade-off between learning and memorization in DCNNs when faced with imperfect training data.

Main Methods:

• Simulated a training dataset with varying proportions of corrupted data and noisy labels to mimic real-world limitations.
• Employed transfer learning strategies for DCNN model training on mammogram data.
• Evaluated DCNN generalization error using the area under the receiver operating characteristic curve (AUC) for both training and test sets.
• Monitored DCNN weight changes post-transfer learning to assess model adaptation and potential overfitting.

Main Results:

• The proportion of corrupted data in the training set significantly influenced the balance between DCNN learning and memorization.
• Higher proportions of noisy data led to increased generalization error, indicating a detrimental effect on model performance.
• Analysis of AUC and weight changes provided insights into how DCNNs adapt to imperfect datasets.

Conclusions:

• Transfer learning strategies for DCNNs in medical imaging, particularly for mammogram classification, must be carefully designed.
• Consideration of training set size and data quality is crucial to minimize memorization and enhance model generalizability.
• Properly designed transfer learning can improve the reliability of AI tools for breast cancer screening.