Deep transfer learning-based prostate cancer classification using 3 Tesla multi-parametric MRI.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer (PCa) detection and classification are crucial for patient management.
• Accurate differentiation between indolent and clinically significant PCa on MRI is challenging.
• Current methods like PIRADS v2 scoring have limitations in distinguishing PCa aggressiveness.

Purpose of the Study:

• To develop and evaluate a deep transfer learning (DTL) model for differentiating indolent from clinically significant prostate cancer (PCa) lesions.
• To compare the performance of the DTL model against a conventional deep learning (DL) model and the PIRADS v2 score.
• To validate the models using whole-mount histopathology (WMHP) on 3 Tesla multi-parametric MRI (3T mp-MRI).

Main Methods:

• A cohort of 140 patients with 3T mp-MRI and WMHP was studied.
• A DTL-based model was trained on 169 lesions from 110 patients and tested on 47 lesions from 30 patients.
• Performance was assessed using accuracy, sensitivity, specificity, and area under the curve (AUC), with comparisons to a DL model without transfer learning and PIRADS v2 score (≥4).

Main Results:

• The DTL-based model achieved an AUC of 0.726, outperforming the DL model (AUC 0.687) and PIRADS v2 score (AUC 0.711) in discriminating clinically significant PCa lesions.
• The DTL model demonstrated superior performance in identifying significant PCa lesions in the testing set.
• The DTL model's AUC was comparable to the PIRADS v2 score (p=0.89).

Conclusions:

• The proposed DTL-based model shows promise for improving the classification of prostate cancer lesions on 3T mp-MRI.
• DTL offers a potential advancement over standard DL techniques and PIRADS v2 scoring for PCa aggressiveness assessment.
• Further validation is warranted to integrate this DTL model into routine clinical practice for enhanced PCa diagnosis.