Multi-task multi-scale learning for outcome prediction in 3D PET images.

Area of Science:

• Oncology
• Radiology
• Medical Imaging
• Machine Learning

Background:

• Precision medicine in oncology relies on predicting patient treatment response and survival.
• Radiomics offers a non-invasive method using medical images for analysis.
• Automated deep learning tools for lesion segmentation are data-intensive, facing challenges with limited annotated medical data.

Purpose of the Study:

• To develop a novel multi-task, multi-scale learning framework for enhanced radiomic analysis.
• To improve the prediction of patient survival and treatment response in oncology.
• To address the data limitations in medical image analysis through a generalized model.

Main Methods:

• Proposed a multi-task, multi-scale learning framework for radiomic analysis.
• Utilized subsidiary tasks to act as an inductive bias for better model generalization.
• Employed an encoder designed to extract powerful features by leveraging multiple tasks.

Main Results:

• The framework was validated for treatment response and survival prediction in esophageal and lung cancers.
• Achieved an area under the ROC curve of 77% for esophageal cancer and 71% for lung cancer.
• Demonstrated superior performance compared to single-task learning methods.

Conclusions:

• Multi-task, multi-scale learning significantly enhances radiomic analysis performance.
• The framework effectively extracts rich information from intratumoral and peritumoral regions.
• This approach advances precision medicine by improving non-invasive patient outcome prediction.