Deep Learning Renal Segmentation for Fully Automated Radiation Dose Estimation in Unsealed Source Therapy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Convolutional Neural Networks (CNNs) excel at object detection and can be trained on limited data.
• Accurate organ recognition in medical imaging enables new quantitative diagnostic techniques.
• CNNs can streamline radiation dose analysis for targeted radionuclide therapies.

Purpose of the Study:

• Develop an automated image segmentation tool using 3D CNNs for kidney contour detection.
• Evaluate the accuracy and efficiency of automated kidney segmentation compared to manual contouring.
• Assess the impact of automated segmentation on renal radiation dose estimation in patients undergoing radionuclide therapy.

Main Methods:

• A 3D CNN model was developed for segmenting right and left kidney contours on non-contrast CT images.
• The model was trained on 89 manually contoured cases and tested on patients receiving 177Lu-PSMA-617 therapy.
• Automated contours were compared to expert-drawn contours using Dice score, mean distance-to-agreement, and volume. Renal radiation dose was estimated using automated contours on SPECT imaging.

Main Results:

• The CNN segmentation accurately identified kidneys in all patients, with mean Dice scores of 0.91 (right) and 0.86 (left).
• The automated system provided contours in approximately 90 seconds, significantly faster than manual methods.
• No significant difference in estimated renal radiation absorbed dose was found between automated and manual contouring, except in cases with cystic kidneys.

Conclusions:

• Automated kidney contouring using CNNs demonstrates high accuracy and efficiency for quantitative assessment of SPECT/PET images.
• This approach shows comparable accuracy to human observers for radiation dose interpretation in unsealed source therapy.
• CNN-based segmentation holds promise for improving quantitative analysis and treatment planning in radionuclide therapy.