Deep Learning Features Can Improve Radiomics-Based Prostate Cancer Aggressiveness Prediction
Nuno M Rodrigues1,2, José Guilherme de Almeida2, Ana Rodrigues2,3
1LASIGE, Department of Informatics, Faculty of Sciences, University of Lisbon, Lisbon, Portugal.
JCO Clinical Cancer Informatics
|September 18, 2024
Summary
Artificial intelligence enhances prostate cancer aggressiveness prediction by integrating deep learning features with radiomic signatures. Careful assessment of deep features is crucial for improving model performance and clinical decision-making.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Artificial intelligence (AI) shows promise in improving prostate cancer detection and treatment.
- Radiomics, traditionally viewed as a sequential process, can be integrated with AI for enhanced analysis.
- Prostate segmentation and reconstruction models offer valuable feature information beyond traditional radiomics.
Purpose of the Study:
- To investigate the impact of integrating deep learning features with radiomic signatures for prostate cancer aggressiveness classification.
- To evaluate the effectiveness of different modeling decisions, including feature aggregation and dimensionality reduction, on predictive performance.
- To determine if combining deep features from segmentation/reconstruction models with radiomic features improves classification accuracy.
Main Methods:
- Conducted 2,244 experiments using deep learning features from 13 different models.
- Extracted deep features from models trained on various prostate anatomic zones.
- Employed feature aggregation and dimensionality reduction techniques, including principal component analysis (PCA).
Main Results:
- Integrating deep features from autoencoder models trained on the full prostate gland with radiomic features significantly improved disease aggressiveness classification.
- While some deep features enhanced prediction, others were detrimental, highlighting the need for careful selection.
- Principal Component Analysis (PCA) and PCA + relief were identified as effective feature selection methods.
Conclusions:
- Combining deep features from prostate reconstruction models with radiomic features can significantly boost the performance of models predicting disease aggressiveness.
- The strategic selection of features is critical for optimizing predictive model performance.
- While deep features offer potential benefits, their inclusion requires thorough evaluation to avoid negatively impacting predictive accuracy.
