Autosegmentation of lung computed tomography datasets using deep learning U-Net architecture.

Area of Science:

• Medical Imaging
• Radiotherapy
• Artificial Intelligence

Background:

• Radiotherapy planning requires extensive imaging data and manual segmentation of organs at risk (OARs), consuming significant clinician time.
• Automating OAR segmentation can streamline radiotherapy workflows and improve treatment precision.

Purpose of the Study:

• To develop and evaluate a U-net-based deep learning architecture for automated segmentation of OARs in lung cancer radiotherapy.
• To assess the accuracy of the U-net model in segmenting the lungs, heart, and spinal cord using Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD).

Main Methods:

• Four U-Net models were trained on CT datasets from 20 lung cancer patients for 100 epochs.
• The models were specifically trained to segment the right lung, left lung, heart, and spinal cord.
• Segmentation accuracy was quantified using DSC and HD metrics, comparing model predictions to manual contours.

Main Results:

• The U-net models achieved high average DSC scores: 0.96 for the left lung, 0.94 for the right lung, and 0.88 for the heart.
• The spinal cord model achieved a DSC of 0.76, with corresponding Hausdorff Distances ranging from 2.76 mm to 4.09 mm.
• Autosegmentation for lungs and heart showed good agreement with manual contours, though the heart segmentation had minor inaccuracies in some cases.

Conclusions:

• U-net-based deep learning models show promise for accurate and efficient autosegmentation of OARs in lung cancer radiotherapy.
• While lung and heart segmentation accuracy is high, further refinement may be needed for structures like the spinal cord due to its size.
• This ongoing research aims to provide a valuable tool for radiation oncologists, reducing manual segmentation effort and potentially improving treatment planning.