Optimizing Inference Distribution for Efficient Kidney Tumor Segmentation Using a UNet-PWP Deep-Learning Model with...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Kidney tumors pose diagnostic challenges due to asymptomatic presentation, necessitating early detection.
• Neural networks show promise for disease prediction but often face computational limitations in clinical settings.

Purpose of the Study:

• Introduce the UNet-PWP architecture for efficient kidney tumor segmentation.
• Address computational complexity constraints of neural networks in medical imaging.
• Enhance model performance and interpretability for clinical application.

Main Methods:

• Developed UNet-PWP architecture with adaptive partitioning to create smaller submodels.
• Augmented UNet depth using pre-trained weights for improved segmentation capabilities.
• Applied weight-pruning techniques to optimize the model and incorporated attention and Grad-CAM XAI for interpretability.

Main Results:

• UNet-PWP achieved 97.01% accuracy on training and test datasets, surpassing DeepLab V3+.
• Adaptive partitioning and weight pruning streamlined the model without performance degradation.
• Explainable AI methods provided insights into model predictions, enhancing clinical trust.

Conclusions:

• The UNet-PWP architecture offers an efficient and accurate solution for kidney tumor segmentation.
• The model's interpretability is crucial for its adoption by healthcare professionals.
• This approach demonstrates the potential of optimized neural networks in clinical diagnostics.