Impact of image compression on deep learning-based mammogram classification.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Medical imaging generates large datasets, necessitating image compression to manage storage and transmission overhead.
• Deep learning models show promise in analyzing medical images for disease detection, but their performance with compressed data requires evaluation.

Purpose of the Study:

• To assess the impact of varying image compression ratios on the performance of deep learning models for mammogram classification.
• To determine if training deep learning models on compressed mammograms affects their diagnostic accuracy and reliability.

Main Methods:

• A retrospective study analyzed 9111 mammograms (normal, benign, malignant) from the National Cancer Center, Republic of Korea.
• Convolutional neural networks (CNNs) were trained on mammograms compressed at ratios from 15:1 to 11:1.
• Model performance was evaluated using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) with five-fold cross-validation.

Main Results:

• Models trained on images with compression ratios (CRs) up to 5:1 achieved an average AUROC of 0.87 and AUPRC of 0.75.
• Performance decreased significantly with higher CRs (10:1 and 11:1), yielding average AUROC of 0.79 and AUPRC of 0.49.
• Saliency maps indicated that models trained on less compressed images better identified diagnostically relevant areas compared to those trained on highly compressed images.

Conclusions:

• Moderate image compression (CR ≤ 5:1) has a limited impact on the performance of deep learning models for mammogram classification.
• High compression ratios (CR > 5:1) can degrade model performance and lead to misinterpretation of image features.
• Training deep learning models on compressed data can improve their robustness when tested on unseen compressed images.