[Paper Review] A Comprehensive Study of Data Augmentation Strategies for Prostate Cancer Detection in Diffusion-weighted MRI using Convolutional Neural Networks

Ruqian Hao, Khashayar Namdar+3 more|arXiv (Cornell University)|Jun 1, 2020

Radiomics and Machine Learning in Medical Imaging25 references4 citations

TL;DR

This study evaluates five data augmentation techniques—random rotation, horizontal flip, vertical flip, random crop, and translation—on diffusion-weighted MRI for prostate cancer detection using 2D convolutional neural networks. The rotation augmentation yielded the highest AUC of 0.85 with a shallow CNN, outperforming deeper architectures and other augmentation strategies on a test set of 95 patients.

ABSTRACT

Data augmentation refers to a group of techniques whose goal is to battle limited amount of available data to improve model generalization and push sample distribution toward the true distribution. While different augmentation strategies and their combinations have been investigated for various computer vision tasks in the context of deep learning, a specific work in the domain of medical imaging is rare and to the best of our knowledge, there has been no dedicated work on exploring the effects of various augmentation methods on the performance of deep learning models in prostate cancer detection. In this work, we have statically applied five most frequently used augmentation techniques (random rotation, horizontal flip, vertical flip, random crop, and translation) to prostate Diffusion-weighted Magnetic Resonance Imaging training dataset of 217 patients separately and evaluated the effect of each method on the accuracy of prostate cancer detection. The augmentation algorithms were applied independently to each data channel and a shallow as well as a deep Convolutional Neural Network (CNN) were trained on the five augmented sets separately. We used Area Under Receiver Operating Characteristic (ROC) curve (AUC) to evaluate the performance of the trained CNNs on a separate test set of 95 patients, using a validation set of 102 patients for finetuning. The shallow network outperformed the deep network with the best 2D slice-based AUC of 0.85 obtained by the rotation method.

Motivation & Objective

• To investigate the impact of various data augmentation strategies on deep learning performance in prostate cancer detection from diffusion-weighted MRI.
• To compare the effectiveness of common augmentation techniques—rotation, flip, crop, and translation—on model generalization and accuracy.
• To assess whether augmentation improves performance in low-data regimes typical in medical imaging.
• To evaluate both shallow and deep CNN architectures under different augmentation schemes.
• To determine the optimal augmentation strategy for maximizing AUC in a clinical MRI classification task.

Proposed method

• Applied five standard data augmentation techniques—random rotation, horizontal flip, vertical flip, random crop, and translation—individually to a DW-MRI dataset of 217 patients.
• Applied augmentations independently to each data channel to preserve multi-parametric MRI characteristics.
• Trained both a shallow and a deep Convolutional Neural Network (CNN) on each of the five augmented datasets separately.
• Used a validation set of 102 patients for hyperparameter tuning and model selection.
• Evaluated final models on a held-out test set of 95 patients using the Area Under the ROC Curve (AUC) as the primary metric.
• Ensured consistent training and evaluation protocols across all augmentation and network configurations.

Experimental results

Research questions

• RQ1 Which data augmentation technique yields the highest classification performance in prostate cancer detection from DW-MRI?
• RQ2 How do different augmentation strategies affect the generalization and robustness of CNNs in low-data medical imaging settings?
• RQ3 Does the performance of shallow CNNs surpass that of deeper architectures when combined with data augmentation in this task?
• RQ4 How do individual augmentation methods compare in terms of AUC improvement over the baseline model?
• RQ5 Is there a synergistic effect between augmentation and network depth in this specific medical imaging application?

Key findings

• The rotation augmentation strategy achieved the highest 2D slice-based AUC of 0.85 on the test set, outperforming all other augmentation techniques.
• The shallow CNN architecture achieved better performance than the deeper network, with the best AUC of 0.85 obtained using rotation augmentation.
• Horizontal and vertical flips yielded lower AUC values compared to rotation, indicating limited benefit in preserving diagnostic features in DW-MRI.
• Random crop and translation augmentations showed moderate performance but did not surpass rotation in effectiveness.
• The combination of rotation augmentation with a shallow CNN provided the most robust and accurate classification performance.
• The study demonstrates that simple, static augmentation strategies can significantly improve model performance in low-data medical imaging scenarios.