Novel real-time tumor-contouring method using deep learning to prevent mistracking in X-ray fluoroscopy

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate tumor tracking is crucial for effective cancer treatment, especially without fiducial markers.
• X-ray fluoroscopy can be hindered by high-density structures like bone, leading to inaccurate tumor tracking.
• Controlling "importance recognition"—prioritizing soft tissue over bone—is key for robust tumor tracking.

Purpose of the Study:

• To develop a novel real-time tumor-contouring method using deep learning with "importance recognition" control.
• To improve the robustness of tumor tracking in the presence of interfering structures like bone in X-ray fluoroscopy.

Main Methods:

• A supervised deep learning approach combined with a novel random overlay method was employed.
• The method trains the model to differentiate between important (soft tissue) and unimportant (bone) structures for contouring.
• Image segmentation using deep learning was utilized for tumor contouring.

Main Results:

• The proposed method achieved accurate tracking of a low-visibility tumor with approximately 1 mm error in a simulated fluoroscopy model.
• A high Jaccard index of approximately 0.95 demonstrated strong similarity between segmented and ground truth tumor regions.
• The method exhibited a fast processing time of 25 ms, enabling real-time application.

Conclusions:

• The developed deep learning method shows feasibility for robust, real-time tumor contouring using fluoroscopy.
• The "importance recognition" strategy effectively mitigates interference from bone structures.
• Further clinical validation using actual fluoroscopy systems is warranted.