Main Session

Sep 29

PQA 06 - Genitourinary Cancer, Gynecological Cancer, and Health Care Access and Engagement

3318 - A Radiomics Based Approach for Predicting Prostate Cancer in PIRADS 3 Lesions on Multi-Parametric MRI

02:15pm - 03:30pm ET

Poster Hall - Exhibit Hall A

Screen: 13

POSTER

Presenter(s)

Rikhil Makwana, BA - Chicago Medical School at Rosalind Franklin University, North Chicago, IL

R. D. Makwana¹, E. Stevenson¹, O. Esengur¹, D. Wunsch¹, S. Harmon², S. Gurram³, P. A. Pinto³, and B. Turkbey¹; ¹National Cancer Institute, Artificial Intelligence Resource, Bethesda, MD, ²Molecular Imaging Program, National Cancer Institute, National Institutes of Health, Bethesda, MD, ³Urologic Oncology Branch, National Cancer Institute, NIH, Bethesda, MD

Purpose/Objective(s):

Prostate image reporting and data system (PIRADS) 3 lesions are considered indeterminate. The decision to biopsy depends on individual factors as opposed to objective guidelines. Herein we proposed both a radiomics and combined radiomics and clinical data-based approach to determine the risk of clinically significant prostate cancer (csPCa) (ISUP>1) using multi-parametric MRI (mpMRI) in this cohort.

Materials/Methods:

We hypothesized that radiomics-based feature analysis could be used to predict the presence of csPCa from mpMRI. We additionally hypothesized that age at biopsy (age), prior biopsy, prior positive biopsy, prostatic specific antigen density (PSAD), and zone of the lesion would also contribute to the model’s predictive power.

Prospectively collected mpMRI scans obtained between 2015 and 2025 at a single quaternary care center showing index PIRADS 3 lesions in patients who underwent targeted prostate biopsy were included in analysis. 5-fold training/validation (80%) splits and a testing (20%) splits were generated at the patient level by the presence of csPCa. Radiomic features were extracted from all scans using PyRadiomics.

The Scikit-Learn python package was used to train a stacking classifier, that combined a Gradient Boosting classifier, Bagging classifier, and AdaBoost, using radiomic features (M1) or radiomic features and clinical data (M2). ElasticNet was used for feature selection.

Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) were reported for csPCa detection with 95% confidence intervals (CI). F1 score and area under the receiver operating curve (AUC), determined through 500 replicate bootstraps, are reported without CI.

Results:

The cohort consisted of 322 consecutive patients (n = 77/401 lesion level csPCa). The training/validation splits included 327 lesions (csPCa = 61). The test set included 74 lesions from 64 patients (csPCa = 16). Median age at biopsy was 66 years (Interquartile Range (IQR): [61, 70]) and PSAD was 0.11 ng/ml² (IQR: [0.08, 0.18]). 71% (286/401) had a previous biopsy of which 71% (204/286) were positive. 64% of lesions were in the transition zone (TZ; 259/401), 33% in the peripheral zone (PZ; 134/401), and 2% in the central zone (7/401).

While M1 and M2 had the same sensitivity 0.5 [CI: 0.28, 0.72], M2 performed better across all other metrics (Table 1).

Conclusion:

The inclusion of clinical data improved the performance of a radiomics based classifier for csPCa within PI-RADS 3 lesions. Additional improvements to model sensitivity and PPV are necessary prior to clinical implementation.

**Table 1: Overview of Model Performance**
Model	M1	M2
Sensitivity [CI]	0.5 [ 0.28, 0.72]	0.5 [ 0.28, 0.72]
Specificity [CI]	0.78 [0.65, 0.86]	0.86 [0.75, 0.93]
Accuracy [CI]	0.72 [0.60, 0.81]	0.78 [0.68, 0.86]
PPV [CI]	0.38 [0.21, 0.60]	0.5 [0.28, 0.72]
NPV [CI]	0.85 [0.73, 0.92]	0.86 [0.75, 0.93]
F1	0.43	0.5
AUC	0.67	0.75