Improving prototypical parts abstraction for case-based reasoning explanations designed for the kidney stone type...

Area of Science:

• Urology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate in-vivo identification of kidney stone types during ureteroscopy is crucial for efficient treatment and preventing recurrence.
• Current visual recognition by urologists is operator-dependent and time-consuming.
• Existing deep learning (DL) models lack transparency, failing to correlate visual features with established morphoconstitutional analysis (MCA) criteria.

Purpose of the Study:

• To develop an interpretable deep learning (DL) model for in-vivo kidney stone type identification during ureteroscopy.
• To ensure the model's decision-making process aligns with visual features used in biological morphoconstitutional analysis (MCA).
• To improve classification accuracy and provide explainable insights for clinical application.

Main Methods:

• A case-based reasoning DL model utilizing prototypical parts (PPs) was developed.
• PPs encode visual features (hue, saturation, intensity, texture) relevant to MCA.
• A novel loss function optimized PP generation, and local/global descriptors explained classification decisions.

Main Results:

• The proposed DL model achieved an overall average classification accuracy of 90.37±0.6% on six common kidney stone types.
• The model demonstrated enhanced explainability by linking visual features to MCA criteria.
• Accuracy slightly surpassed the best existing DL models (88.2±2.1%) while offering superior interpretability.

Conclusions:

• The interpretable DL model offers a significant advancement for in-vivo kidney stone identification in urology.
• The model's ability to explain its decisions fosters trust and facilitates clinical adoption of AI solutions.
• This approach bridges the gap between AI-driven analysis and traditional biological assessment of kidney stones.