Predicting spatial esophageal changes in a multimodal longitudinal imaging study via a convolutional recurrent neural...

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Acute esophagitis (AE) is a common complication in lung cancer patients treated with radiotherapy.
• Early prediction of AE is crucial for reducing esophageal toxicity and enabling adaptive radiotherapy (ART).

Purpose of the Study:

• To develop and validate a novel machine learning framework for predicting patient-specific esophageal changes following radiotherapy.
• To facilitate early detection of AE and support ART decision-making.

Main Methods:

• A machine learning framework combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs) was developed.
• The algorithm analyzes MRI/CBCT scans to predict esophageal spatial presentation and volume changes over time.
• Trained and validated on institutional data and externally validated using a public dataset.

Main Results:

• The algorithm achieved a Dice coefficient of 0.83 ± 0.04 for predicting esophageal condition and a high correlation (0.98) for volume estimation using institutional data.
• External validation using CBCT data yielded an average Dice coefficient of 0.89 ± 0.03.
• Early prediction of esophageal changes over 3 weeks was demonstrated using the first three weekly scans.

Conclusions:

• The developed machine learning framework shows promise for early monitoring and detection of AE in radiotherapy patients.
• This tool could significantly contribute to the ART decision-making process by enabling proactive management of esophageal toxicity.