Modeling Texture in Deep 3D CNN for Survival Analysis.

Area of Science:

• Radiomics and Medical Imaging Analysis
• Cancer Prognostics
• Machine Learning in Healthcare

Background:

• Radiomics shows promise for predicting pancreatic ductal adenocarcinoma (PDAC) survival.
• Convolutional neural networks (CNNs) are underutilized in radiomics due to large training data requirements.

Purpose of the Study:

• To propose a novel radiomic descriptor, Gaussian mixture model-CNN (GMM-CNN), to enhance PDAC survival prediction.
• To evaluate the efficacy of GMM-CNN features against existing methods for pre-operative prognostication.

Main Methods:

• GMM-CNN features were computed from pre-operative CT scans of PDAC patients.
• A random forest (RF) classifier utilized GMM-CNN features for survival outcome prediction.
• Performance was compared against 3D CNN output, standard radiomics, conditional entropy (CENT), and clinical variables.

Main Results:

• GMM-CNN features achieved the highest AUC of 72.0% (p < 0.0001) using the RF model.
• This outperformed 3D CNN (64.0%), standard radiomics (66.8%), CENT (64.2%), and clinical features (57.6%).

Conclusions:

• The proposed GMM-CNN features significantly enhance the ability to predict PDAC patient survival.
• This method offers a powerful tool for pre-operative prognostication using routine imaging data.