利用深度學習模型結合非侵入性的核磁共振水脂比與攝護腺特異抗原數值預測攝護腺癌的侵襲性

謝宜岑¹、郭永明²、陳嘉豪²、楊峻泓³、謝嘉興⁴

¹國立成功大學醫學院附設醫院 教學中心；²國立虎尾科技大學 電子工程系，³電機工程系；⁴衛生福利部臺南醫院 泌尿科

Deep learning-based artificial intelligence model for predicting prostate cancer aggressiveness using non-invasive water-fat ratio in magnetic resonance imaging and prostate-specific antigen

Yi-Tsen Hsieh¹, Yung-Ming Kuo², Jia-Hao Chen², Chun-Hung Yang³, Gia-Shing Shieh⁴

¹Education Center, National Cheng Kung University Hospital, College of Medicine, National Cheng Kung University, Tainan, Taiwan; ²Department of Electronic Engineering and ³Department of Electrical Engineering, National Formosa University, Yunlin, Taiwan; ⁴Department of Urology, Tainan Hospital, Ministry of Health and Welfare, Executive Yuan, Tainan, Taiwan

Purpose: Pretreatment assessment in prostate cancer holds several clinical limitations. Currently, the PI-RADS scoring in magnetic resonance imaging (MRI) exhibits diagnostic uncertainty when distinguishing low-risk lesions from clinically significant prostate cancer. While conventional diagnostic methods can be invasive and limited in efficacy, the development of AI-driven pelvic MRI analysis offers potential. However, few studies have prospectively validated these tools for pretreatment detection of prostate cancer. We developed a non-invasive deep learning tool combining prostate-specific antigen (PSA) levels with enhanced T1-weighted (eT1w) images, enhanced T2-weighted (eT2w) images, and the novel Imitation Water–Fat Ratio (I-WFR) from MRI, for early detection and risk assessment of prostate cancer.

Materials and Methods: This single-center, prospective study enrolled 45 patients between January 2021 and December 2023. Patients were categorized into low-risk (ISUP grade 1 and PSA < 10 ng/mL and cT1-2a) and non–low-risk groups. Pelvic MRI images from patients with PI-RADS 4 and 5 were included in the study. This study constructed a multimodal deep learning model utilizing patients' eT1w images, eT2w images, and PSA data as primary inputs, and the Region of Interest (ROI) and the I-WFR were obtained via deep learning methods. The I-WFR was calculated from the tumor-surrounding fat to detect brown adipose tissue via water and fat content, which has been shown to be correlated with prostate cancer progression. Corresponding data and image preprocessing steps were executed independently to ensure the quality and consistency of the input data. ResNet34-FCN models utilizing only PSA levels (ResNet34-FCN-PSA), MRI data features (ResNet34-FCN-DF), or both data inputs (ResNet34-FCN-ALL) were evaluated to compare their diagnostic performance in the risk assessing of prostate cancer. To identify the optimal model, we utilized several performance metrics including Accuracy, F1-Score, Sensitivity, Specificity, and AUC, along with the Youden’s Index (J = Sensitivity + Specificity - 1) to achieve a comprehensive evaluation of model performance.

Results: Resnet34-FCN-PSA achieved an accuracy of 0.488, an AUC of 0.452, a J-statistic of –0.094, an F1 score of 0.303, a sensitivity of 0.357, and a specificity of 0.548. Resnet34-FCN-DF reached an accuracy of 0.511, an AUC of 0.468, a J-statistic of –0.060, an F1 score of 0.312, a sensitivity of 0.357, and a specificity of 0.580. Furthermore, Resnet34-FCN-ALL demonstrated an accuracy of 0.822, an AUC of 0.792, a J-statistic of 0.585, an F1 score of 0.714, a sensitivity of 0.714, and a specificity of 0.870.

Conclusion: ResNet34-FCN-ALL, which incorporates PSA and MRI data, provides a non-invasive tool for prostate cancer assessment by using the I-WFR as a novel imaging biomarker. The model improves diagnostic accuracy and patient safety by minimizing invasive examinations in low-risk patients. In conclusion, ResNet34-FCN-ALL represents a valuable tool for integration into clinical prostate cancer management.