Related Experiment Video
Updated: May 5, 2026
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
42.8K
Impact of Downsampling Size and Interpretation Methods on Diagnostic Accuracy in Deep Learning Model for Breast
Ryusei Inamori1, Tomofumi Kaneno1, Ken Oba2
1Tohoku University Graduate School of Medicine, Department of Clinical Imaging.
The Tohoku Journal of Experimental Medicine
|July 24, 2024
Summary
Image downsampling and interpolation methods significantly impact deep learning model accuracy for breast cancer diagnosis using digital breast tomosynthesis (DBT). Careful selection is crucial for reliable diagnostic performance.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Deep learning (DL) models show potential for breast cancer diagnosis with digital breast tomosynthesis (DBT).
- The influence of varying image matrix sizes and interpolation techniques on DL diagnostic accuracy in DBT is not well understood.
- Optimizing preprocessing is key for efficient DL model training and improved diagnostic outcomes.
Purpose of the Study:
- To investigate the impact of different matrix sizes and image interpolation methods on the diagnostic accuracy of DL models for breast cancer detection using DBT images.
- To identify optimal preprocessing strategies for enhancing DL model performance in DBT analysis.
Main Methods:
- A retrospective study included 499 patients who underwent DBT.
- Images were downsampled to various matrix sizes (256x256, 128x128, 64x64, 32x32) using five interpolation methods: Nearest (NN), Bilinear (BL), Bicubic (BC), Hamming (HM), and Lanczos (LC).
- Diagnostic accuracy was evaluated using the area under the curve (AUC) with 95% confidence intervals (CI).
Main Results:
- Downsampling to 256x256 with Lanczos (LC) and 128x128 with Hamming (HM) and LC interpolation significantly reduced AUC compared to original 512x512 images.
- All interpolation methods resulted in a significant decrease in AUC for 64x64 and 32x32 downsampled images.
- The choice of downsampling size and interpolation method critically affects DL model diagnostic performance.
Conclusions:
- Image downsampling matrix size and interpolation methods significantly influence the diagnostic accuracy of DL models in DBT.
- Understanding these effects is vital for optimizing preprocessing steps to improve the reliability and accuracy of AI-driven breast cancer diagnosis.
- Further research into optimal preprocessing parameters can enhance computational efficiency and clinical utility of DL in medical imaging.
