Deep Semisupervised Transfer Learning for Fully Automated Whole-Body Tumor Quantification and Prognosis of Cancer on...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate cancer detection and characterization are crucial for timely and effective treatment.
• Whole-body tumor segmentation and prognosis using PET/CT imaging present significant clinical challenges.

Purpose of the Study:

• To develop a deep, semisupervised transfer learning approach for automated whole-body tumor segmentation and prognosis on PET/CT scans.
• To evaluate the performance of this approach across various cancer types, including lung, melanoma, lymphoma, head and neck, breast, and prostate cancers.

Main Methods:

• A retrospective study utilized 611 18F-FDG PET/CT scans and 408 PSMA PET/CT scans.
• A nnU-net backbone was employed for segmentation using limited annotations and radiomics analysis.
• Prognostic models were developed for risk stratification, survival estimation, and treatment response prediction.

Main Results:

• The approach achieved high median Dice similarity coefficients for tumor segmentation across different cancers (e.g., 0.81 for lung cancer, 0.83 for lymphoma).
• Prognostic models demonstrated strong performance, with an AUC of 0.86 for prostate cancer risk stratification and significant associations with survival for head and neck cancer.
• Predictive models for breast cancer showed accuracies of 0.72 and 0.84 for predicting pathologic complete response.

Conclusions:

• The developed deep learning approach enables accurate, automated tumor segmentation on PET/CT scans.
• The method shows significant potential for cancer prognosis, including risk stratification and treatment response prediction across multiple cancer types.
• This automated approach can optimize early cancer treatment by providing reliable imaging insights.