Adaptive Sparsity Regularization Based Collaborative Clustering for Cancer Prognosis.

Area of Science:

• Medical Imaging
• Radiomics
• Machine Learning

Background:

• Radiomic approaches show promise in predicting cancer patient outcomes.
• Feature dimensionality reduction is crucial but faces challenges with data noise and latent supervision.

Purpose of the Study:

• To develop a robust feature dimensionality reduction method for radiomics.
• To improve patient stratification and clinical outcome prediction accuracy.

Main Methods:

• An adaptive sparsity regularization-based collaborative clustering method was developed.
• This method uses adaptive sparsity regularized matrix tri-factorization for simultaneous denoising and dimension reduction.
• Bayesian framework for sparsity regularization grounded on distribution modeling.

Main Results:

• The proposed method effectively clusters data and outperforms alternatives on synthetic data.
• Experiments on FDG-PET/CT rectal cancer data showed superior patient stratification.
• Improved prediction of patient clinical outcomes was demonstrated.

Conclusions:

• The developed method offers improved discriminative power and robustness for radiomic features.
• This approach enhances the utility of radiomics in clinical outcome prediction and patient stratification.