Convolutional neural network-based model observer for signal known statistically task in breast tomosynthesis images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Human observer studies for image quality assessment are resource-intensive.
• Existing mathematical model observers often assume exact signal knowledge, limiting real-world applicability.
• This limitation is particularly relevant in tasks involving unknown signal size and shape.

Purpose of the Study:

• To propose a convolutional neural network (CNN)-based model observer for breast tomosynthesis.
• To evaluate the model for signal known statistically (SKS) and background known statistically (BKS) detection tasks.
• To address limitations of models assuming exact signal information.

Main Methods:

• A CNN-based model observer was developed and trained on breast tomosynthesis images.
• Acquisition parameters included six different angles (10°-60°) at 2.3 mGy dose.
• Performance was compared against the Hotelling observer (HO) using spherical (SKE) and spiculated (SKS) signals; pixel-wise gradient-weighted class activation mapping (pGrad-CAM) was used for visualization.

Main Results:

• The CNN-based model observer demonstrated superior detection performance over the HO across all tasks.
• Performance improvement was more pronounced for SKS tasks, indicating enhanced handling of signal/background variations.
• pGrad-CAM effectively localized discriminative regions, and the CNN model required fewer images for comparable performance.

Conclusions:

• A CNN-based model observer was successfully developed for SKS and BKS detection in breast tomosynthesis.
• The proposed CNN observer significantly outperformed the Hotelling observer.
• This approach offers a more robust and efficient method for image quality assessment in mammography.