Weakly Supervised Tumor Detection in PET Using Class Response for Treatment Outcome Prediction.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Radiomics

Background:

• Radiomic features from tumors predict outcomes, but manual segmentation is time-consuming and requires expertise.
• Supervised deep learning for medical image segmentation needs large datasets, which are often unavailable.
• Automating tumor segmentation is crucial for efficient and accurate radiomic analysis.

Purpose of the Study:

• To develop a weakly supervised deep learning framework for 3D tumor segmentation in PET images.
• To enable accurate outcome prediction using segmented tumors with limited data and weak ground truth.
• To automate the time-consuming process of lesion segmentation for radiomic analysis.

Main Methods:

• A 3D tumor segmentation in PET images using a weakly supervised deep learning method.
• Calculation of maximum intensity projection (MIP) images from 3D PET data.
• Classification of MIP images and generation of class activation maps via multitask learning.
• 3D tumor segmentation using 2D activation maps and a novel multitask loss function.
• Outcome prediction using a 3D-CNN classifier on segmented tumor regions.

Main Results:

• The proposed approach achieves state-of-the-art prediction results with small datasets and weak segmentation ground truth.
• Validated for treatment response and survival in lung and esophageal cancers.
• Achieved an AUC of 67% for lung cancer and 59% for esophageal cancer.
• Obtained a dice coefficient of 73% for lung cancer and 0.77% for esophageal cancer segmentation.

Conclusions:

• Weakly supervised deep learning can effectively automate 3D tumor segmentation in PET images.
• The framework allows for accurate outcome prediction with limited medical imaging data.
• This method significantly advances computer-aided detection (CAD) for radiomic analysis and cancer outcome prediction.