Deep learning based on multiparametric MRI for differentiating benign from malignant PI-RADS 3 prostate lesions: a...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer diagnosis presents challenges, particularly with indeterminate PI-RADS 3 lesions.
• Multiparametric MRI (mpMRI) is crucial, but differentiating benign from malignant PI-RADS 3 lesions requires advanced tools.

Purpose of the Study:

• To develop and validate a non-invasive deep learning radiomics model for classifying PI-RADS 3 prostate lesions.
• To assess the model's performance against clinical data and identify its utility in clinical decision-making.

Main Methods:

• A retrospective study of 382 patients with PI-RADS 3 lesions was conducted across two centers.
• Deep learning features (ResNet101) and radiomics features (mRMR, LASSO) were extracted from mpMRI (T2WI, DWI, ADC).
• Clinical, deep-learning radiomics signature (DLRS), and combined models were constructed and compared.

Main Results:

• The combined model demonstrated superior performance with high AUCs (0.951, 0.964, 0.925), significantly outperforming the clinical model (p < 0.05).
• The DLRS model also exhibited strong diagnostic capability (AUCs: 0.946, 0.932, 0.918).
• A well-calibrated nomogram and positive decision curve analysis (DCA) indicated clinical utility.

Conclusions:

• The developed mpMRI-based deep learning model effectively differentiates malignant from benign PI-RADS 3 prostate lesions.
• This non-invasive approach offers a valuable tool to support clinical decisions and minimize unnecessary prostate biopsies.