Optimal neural network architecture selection: improvement in computerized detection of microcalcifications

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Microcalcification clusters on mammograms are key indicators of breast cancer.
• Computer-aided detection (CAD) systems aim to improve diagnostic accuracy.
• Convolutional neural networks (CNNs) show promise in enhancing CAD system performance by reducing false positives.

Purpose of the Study:

• To evaluate the impact of optimal neural network architecture selection on CAD system performance for microcalcification detection.
• To compare the accuracy of an automated CNN optimization technique against manual selection for microcalcification detection.

Main Methods:

• A computer program was developed for automatic microcalcification cluster detection on digitized mammograms.
• An automated technique was used to optimize the CNN architecture.
• Performance was evaluated using an independent dataset of 472 mammograms with 253 malignant clusters.

Main Results:

• At a false-positive rate of 0.7 cluster per image, the optimized CNN achieved 84.6% sensitivity, compared to 77.2% for the manually selected CNN (film-based).
• In a case-based approach (detection on either view), the optimized CNN achieved 93.3% sensitivity versus 87.0% for the manually selected CNN at the same false-positive rate.

Conclusions:

• Automated optimization of CNN architecture is effective in improving microcalcification detection accuracy.
• Optimized CNNs significantly reduce false-positive findings, enhancing the reliability of CAD systems for mammography.