Targeting Precision with Data Augmented Samples in Deep Learning.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning (DL) is state-of-the-art for medical image analysis.
• Data augmentation is common for small medical datasets but can overestimate network performance due to correlated samples.
• Current methods often overlook optimizing precision alongside accuracy in augmented datasets.

Purpose of the Study:

• To address the overestimation of network performance caused by correlated samples in data augmentation.
• To propose a novel approach that optimizes both accuracy and precision using augmented data.
• To enhance the robustness and invariance of deep learning models to minor input variations in medical imaging.

Main Methods:

• Developed a new approach leveraging augmented data within a specifically-designed loss function.
• Focused on optimizing precision by considering multiple replicates of the same training data.
• Applied the strategy to two distinct deep learning applications: regression and segmentation in medical imaging.

Main Results:

• The proposed method improves both the overall performance and the network's precision.
• Demonstrated the strategy's effectiveness in both regression and segmentation tasks.
• The approach helps networks become invariant to small variations of the same input samples.

Conclusions:

• Optimizing for precision in data augmentation is crucial for reliable deep learning in medical imaging.
• The proposed loss function effectively utilizes augmented data to improve model robustness.
• This work facilitates the explicit integration of data augmentation within the loss function for enhanced medical image analysis.