運用人工智慧模型自動分割攝護腺癌患者的MRI影像於包膜外侵犯之預測

許程皓¹、賴志慶^4,5、林志杰^1,2,4、黃逸修^1,2,4、鍾孝仁^1,2,4、黃奕燊^1,2,4、魏子鈞^1,2,4
黃子豪^1,2,4、陳威任^1,2,4、張延驊^1,2、彭昱璟^3,4、黃志賢^1,2,4、沈書慧^4,5、范玉華^1,2,4

¹臺北榮民總醫院泌尿部; ²國立陽明交通大學書田泌尿科學研究中心; ³臺北榮民總醫院病理檢驗部
⁴國立陽明交通大學醫學院醫學系; ⁵臺北榮民總醫院放射線部

AI-based Auto-segmentation of MRI for Predicting Extraprostatic Extension
in Prostate Cancer
Chen-Hao Hsu¹, Chih-Ching Lai^4,5, Chih-Chieh Lin^1,2,4, Eric Yi-Hsiu Huang^1,2,4,
Hsiao-Jen Chung^1,2,4, I-Shen Huang^1,2,4, Tzu-Chun Wei^1,2,4, Tzu-Hao Huang^1,2,4, Wei-Jen Chen^1,2,4
Yen-Hwa Chang^1,2, Yu-Ching Peng^3,4, William J. Huang^1,2,4, Shu-Huei Shen^4,5, Yu-Hua Fan^1,2,4

¹ Department of Urology, Taipei Veterans General Hospital, Taipei, Taiwan

² Department of Urology, School of Medicine, College of Medicine and Shu-Tien

Urological Research Center, National Yang Ming Chiao Tung University, Taipei, Taiwan

³ Department of Pathology and Laboratory Medicine, Taipei Veterans General Hospital, Taipei, Taiwan

⁴ School of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan

⁵ Department of Radiology, Taipei Veterans General Hospital, Taipei, Taiwan

Purpose: 

Magnetic resonance imaging (MRI) frequently overestimates extraprostatic extension (EPE) in prostate cancer. We evaluated whether quantitative imaging features derived from an artificial intelligence (AI)-based auto-segmentation model could distinguish pathological T3a disease from overstaged cases.

Materials and methods:  
We retrospectively analyzed patients with MRI-staged T3a prostate cancer who underwent robot-assisted laparoscopic radical prostatectomy between April 2023 and November 2024. After excluding patients without MRI performed at our institution, prior prostate surgery, or hormonal therapy, 171 patients were included. For primary analysis, patients with pathologically confirmed T3a and concordant EPE location (n = 70) were compared with those downstaged to pT2 (n = 75), excluding cases with discordant EPE location. An in-house AI model developed by our radiology team was applied to biparametric MRI to automatically segment the prostate and tumor, generating tumor volume and tumor percentage. Clinical and imaging variables were compared between groups using the Mann–Whitney U test, and the discriminative performance of AI-derived parameters was evaluated using receiver operating characteristic (ROC) curve analysis.

Results:

Pathological staging revealed that 75 (43.9%) of MRI T3a patients were downstaged to pT2, while 88 (51.4%) had pT3a and 8 (4.7%) had pT3b disease. Baseline characteristics including age and prostate volume were comparable between the matched T3a group (pathologically confirmed T3a with concordant EPE, n = 70) and the downstaged pT2 group (n = 75). However, the matched T3a group had significantly higher PSA (14.88 ± 13.63 vs. 9.45 ± 4.83 ng/mL, p = 0.004) and PSA density (0.38 ± 0.36 vs. 0.23 ± 0.14, p < 0.001). Higher-grade disease (pathological Gleason score >= 4+3) was more frequently observed in the matched T3a group. Tumor volume was numerically higher in the matched T3a group compared with the downstaged pT2 group (3.46 ± 3.44 vs. 2.45 ± 1.61 cm³, p = 0.271); tumor percentage was significantly higher in the matched T3a group (0.11 ± 0.10 vs. 0.06 ± 0.04, p = 0.013). ROC curve analysis demonstrated that tumor percentage had an AUC of 0.75 for predicting true EPE. An optimal tumor percentage cutoff value of 0.10 yielded a sensitivity of 67% and a specificity of 74%.