Deep learning tools for the cancer clinic: an open-source framework with head and neck contour validation

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Radiotherapy planning

Background:

• Increasing adoption of deep learning in medicine necessitates clinical staff familiarity.
• Developed an open-source framework for automated head and neck CT contouring using CNNs.

Purpose of the Study:

• To evaluate the performance of a CNN-based automatic contouring framework against a commercial atlas-based tool.
• To provide an accessible deep learning tool for clinical use in radiotherapy planning.

Main Methods:

• Trained separate CNNs for 11 organs at risk using 229 patient CT scans.
• Validated CNN performance against an atlas-based tool using 19 test cases.
• Assessed contour accuracy using quantitative metrics and physician review, including correction time.

Main Results:

• CNNs outperformed atlas-based methods on quantitative metrics for 7 of 11 organs.
• Physician reviews indicated fewer substantial corrections needed for CNN contours.
• CNN contouring required less cumulative correction time than atlas-based methods for most cases.

Conclusions:

• The developed CNN framework is validated for automatic contouring in head and neck radiotherapy.
• An open-source, browser-based interface facilitates easy use and reproducibility.
• This deep learning tool empowers novice personnel with automated contouring capabilities.