Clinical Report: Initial Investigation of Entropy-Weighted Hybrid Pooling in CNNs
Overview
This study introduces an adaptive entropy-weighted hybrid pooling method for CNNs in breast cancer detection using ultrasound imaging. Results from the 3-block CNN indicate an accuracy of 93.98% ± 1.72% and an AUC of 0.9870, while the 4-block CNN achieved an accuracy of 92.90%.
Background
Breast cancer is a leading cause of cancer-related mortality. Ultrasound imaging is a valuable tool for screening due to its non-invasive nature, but it faces challenges such as speckle noise and low contrast. Advances in deep learning, particularly CNNs, offer potential solutions to enhance diagnostic accuracy in ultrasound imaging.
Data Highlights
CNN Architecture
Pooling Method
Accuracy (%)
AUC
3-block
Hybrid
93.98 ± 1.72
0.9870
3-block
Max
92.72 ± 0.85
0.9815
4-block
Hybrid
92.90
N/A
4-block
Max
94.79
N/A
Key Findings
In the 3-block CNN, hybrid pooling achieved an accuracy of 93.98% ± 1.72%, exceeding max pooling at 92.72% ± 0.85%.
In the 4-block CNN, hybrid pooling achieved an accuracy of 92.90%, remaining competitive with max pooling at 94.79%.
Deeper CNN architectures improved accuracy and noise robustness but required careful regularization.
The proposed method adapts pooling based on local image complexity using Shannon entropy.
Hybrid pooling addresses the limitations of conventional max and average pooling methods.
Clinical Implications
Further validation and clinical integration of the adaptive hybrid pooling approach are necessary to assess its effectiveness in real-world settings.
Conclusion
The study presents an adaptive pooling technique that shows potential for ultrasound breast cancer diagnosis. Future research should focus on validating these findings.
