PENet: Prior evidence deep neural network for bladder cancer staging

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Bladder cancer staging is crucial for treatment and prognosis.
• Deep Convolutional Neural Networks (DCNNs) excel at feature extraction but lack interpretability.
• Clinical knowledge of tumor infiltration aids in accurate bladder cancer staging.

Purpose of the Study:

• To introduce PENet, a prior evidence deep neural network, for bladder cancer staging.
• To enhance DCNN performance by incorporating clinical knowledge into the staging process.
• To improve the accuracy and interpretability of bladder cancer staging using MR images.

Main Methods:

• PENet measures tumor penetration of the bladder wall to establish prior evidence.
• Bayesian Theorem is used to formulate the posterior distribution of class probability.
• The loss function is modified based on the posterior distribution, incorporating both prior and prediction evidence.

Main Results:

• PENet demonstrated reduced prediction error and variance when provided with ground-truth consistent prior evidence.
• Experiments showed PENet outperformed standard image-based DCNN algorithms for bladder cancer staging.
• The integration of clinical knowledge improved the alignment of predictions with medical principles.

Conclusions:

• PENet offers a promising approach for accurate and clinically relevant bladder cancer staging.
• Incorporating prior clinical evidence into deep learning models enhances their performance and interpretability.
• This method provides a foundation for developing more reliable AI tools in oncology.