Multi-view radiomics and deep learning modeling for prostate cancer detection based on multi-parametric MRI

Area of Science:

• Radiology
• Medical Imaging
• Oncology

Background:

• Prostate cancer (PCa) and benign prostate hyperplasia (BPH) share similar symptoms, complicating diagnosis.
• Accurate differentiation is crucial for appropriate treatment and avoiding unnecessary biopsies.

Purpose of the Study:

• To develop and compare multi-parametric MR imaging models for distinguishing PCa from BPH.
• To evaluate the performance of radiomics and deep learning approaches, individually and combined.

Main Methods:

• A multi-view radiomics strategy was employed, comparing feature categories (original, LoG, wavelet) and utilizing mRMR and LASSO for feature selection.
• Deep learning (DL) models were constructed using a Swin Transformer architecture with transfer learning.
• Radiomics and DL models were combined, and their predictive performance was evaluated using accuracy, consistency, and clinical benefit metrics.

Main Results:

• The optimal radiomics feature set combined LoG and wavelet features.
• The ADCCombinedScore model, integrating radiomics and DL, exhibited the best predictive performance.
• Adding T2-based models to ADC-based models did not improve, and sometimes reduced, predictive performance.

Conclusions:

• The developed imaging models effectively differentiate PCa from BPH.
• The ADCCombinedScore model, presented as a nomogram, can aid clinicians in treatment decisions and reduce unnecessary biopsies.