Strategy to implement a convolutional neural network based ideal model observer via transfer learning for multi-slice...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computational Imaging

Background:

• Deep learning model observers, particularly convolutional neural networks (CNNs), require extensive training data.
• The number of training samples needed increases with image dimensionality, such as in multi-slice medical imaging.
• Conventional linear model observers often have limitations in complex detection tasks.

Purpose of the Study:

• To develop a CNN-based multi-slice ideal model observer utilizing transfer learning (TL-CNN).
• To reduce the number of training samples required for training deep learning model observers.
• To evaluate the performance and robustness of the TL-CNN compared to conventional observers.

Main Methods:

• Generated simulated breast CT image volumes reconstructed using the Feldkamp-Davis-Kress algorithm.
• Trained and evaluated model observers on background-known-statistically (BKS)/signal-known-exactly and BKS/signal-known-statistically tasks.
• Compared TL-CNN performance against multi-slice channelized Hotelling observer (CHO) and volumetric CHO, analyzing detectability with varying training samples.

Main Results:

• The TL-CNN achieved comparable performance with a 91.7% reduction in training samples compared to non-transfer learning approaches.
• CNN-based observers demonstrated 45% higher detectability in signal-known-statistically tasks and 13% higher in SKE tasks versus linear observers.
• Filter weight correlation analysis confirmed the effectiveness of transfer learning in multi-slice model observer training.

Conclusions:

• Transfer learning significantly reduces the data requirements for training deep learning-based multi-slice model observers.
• The proposed TL-CNN method enhances detection performance in medical imaging tasks without compromising accuracy.
• This approach offers a practical solution for developing robust AI-driven observers in resource-limited scenarios.