Managing tumor changes during radiotherapy using a deep learning model.

Area of Science:

• Radiation Oncology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Lung tumors, particularly non-small-cell lung cancer (NSCLC), often exhibit significant shrinkage during radiation therapy.
• Conventional treatment planning typically does not account for this dynamic tumor volume change, potentially leading to suboptimal dose delivery and increased toxicity.
• Accurate prediction of tumor shrinkage is crucial for adaptive radiotherapy to maintain target coverage and minimize off-target radiation exposure.

Purpose of the Study:

• To develop and evaluate a novel treatment planning framework that incorporates weekly lung tumor shrinkage prediction using deep learning.
• To assess the accuracy of a deep learning model in predicting tumor deformation based on cone beam computed tomography (CBCT) data.
• To compare the dosimetric outcomes of adaptive plans, which account for shrinkage, against conventional plans that do not.

Main Methods:

• A deep learning model was trained to predict weekly tumor deformation using spatial and temporal features from serial CBCT images of 16 NSCLC patients.
• The model predicted tumor contours from Week 3 onwards, based on data from previous weeks, and was validated using Dice Similarity Coefficient (DSC), precision, Average Surface Distance (ASD), and Hausdorff Distance (HD).
• Treatment plans were re-optimized weekly based on predicted tumor contours, aiming to maximize target dose coverage and minimize toxicity to organs at risk (OARs).

Main Results:

• Lung tumors showed an average volume reduction of 38% over six weeks of treatment.
• The deep learning model achieved high prediction accuracy, with DSCs ranging from 0.78 to 0.82 and ASDs from 1.49 to 2.12 mm, significantly outperforming rigid contour transfer.
• Adaptive re-planning maintained target coverage while reducing mean lung dose by an average of 2.85 Gy and also decreased doses to other OARs like the esophagus in some cases.

Conclusions:

• A deep learning-based model effectively predicts lung tumor shrinkage using serial CBCT data, demonstrating high accuracy in contour prediction.
• The proposed adaptive radiotherapy framework successfully maintained target coverage and reduced radiation dose to surrounding healthy tissues, including lungs and esophagus.
• This approach offers a promising strategy for personalized and adaptive radiation therapy in NSCLC, improving treatment efficacy and patient safety.