Explainable Spatial Clustering: Leveraging Spatial Data in Radiation Oncology.

Area of Science:

• Medical physics and data science intersection.
• Application of machine learning in radiation oncology.

Background:

• Increasing data in radiation therapy enables data mining and machine learning.
• Collaboration between machine learning experts and clinicians is crucial for model development and adoption.
• Interpretability of spatial clustering in medical data, particularly for clinical audiences, is under-explored.

Purpose of the Study:

• To reflect on the design of visualizations for explaining novel spatial clustering approaches.
• To address the interpretability of complex anatomical data clustering for clinical users.
• To provide lessons learned for creating visual and explainable spatial clustering methods.

Main Methods:

• Participatory design approach involving radiation oncologists and statisticians.
• Development of visualizations for explaining spatial clustering of head and neck cancer patient data.
• Multi-year collaboration to refine visual explanations.

Main Results:

• Successful co-design of visualizations tailored for clinical interpretation.
• Identification of key challenges and solutions in explaining spatial clustering to clinicians.
• Distillation of practical lessons learned from the collaborative process.

Conclusions:

• Effective visualization is essential for the clinical adoption of spatial clustering in radiation therapy.
• Collaborative, participatory design is vital for developing interpretable machine learning tools for clinicians.
• The study provides a framework for creating explainable AI in medical imaging and spatial data analysis.