Stephanie A. Harmon

Purpose: Microsatellite instability–high/mismatch repair–deficient (MSI-H/dMMR) tumors exhibit distinct biological features and are associated with Lynch syndrome. Current approaches for detecting MSI-H/dMMR status rely on immunohistochemistry or molecular assays. Recent advances in digital pathology have enabled the extraction of clinically relevant molecular tumor features from routine hematoxylin and eosin (H&E)–stained diagnostic slides. Image-based deep learning approaches have the potential to screen for MSI-H/dMMR at scale in tumor types such as localized prostate cancer, in which MSI-H/dMMR is so rare that traditional testing is not routinely performed. Experimental Design: We developed and validated PathStage-MSI, a multistage transfer learning framework to predict MSI-H/dMMR status from H&E-stained slides. The model was trained and validated on 984 standard-of-care whole-slide images (WSI; n = 906 patients) from three international clinical prostate cancer cohorts with matched H&E images and genomic data. To assess model generalizability and stability, we additionally evaluated PathStage-MSI in an independent metastatic prostate cancer cohort comprising 85 WSIs from 29 patients. Results: PathStage-MSI achieved the area under the receiver operating characteristic curve (AUROC) of 0.89 (0.79–0.96) and 0.91 (0.81–0.98) in two internal hold-out validation cohorts and 0.75 (0.57–0.88) and 0.81 (0.66–0.94) in two independent external validation cohorts. Conclusions: PathStage-MSI is a deep learning–based image analysis framework for detecting MSI-H/dMMR prostate cancer directly from routine histopathology. By enabling efficient prescreening for this rare molecular subtype, it has the potential to improve patient identification for precision therapies and appropriate germline genetic testing.

Background: Management of men with equivocal MRI findings remains uncertain, while low‐risk MRI may still harbor clinically significant prostate cancer (csPCa). We examined whether prostate‐specific antigen (PSA)‐based measures could identify a low‐risk subgroup suitable for biopsy deferral in patients with Prostate Imaging–Reporting and Data System (PI‐RADS) 1–3 lesions, and which clinical & imaging factors further refined risk stratification. Methods: This prospective single‐center study included 473 consecutive patients with index PI‐RADS 1–3 findings on prostate MRI who underwent systematic and/or MRI/US fusion‐guided biopsy between May 2019 and December 2025. Patient‐level csPCa (grade group ≥ 2) detection rates (CDR) were calculated across PI‐RADS groups. csPCa detection, biopsy avoidance, and missed csPCa were assessed using biopsy‐deferral thresholds based on PSA (< 3 ng/mL), age‐adjusted PSA, and PSA density (PSAD) (< 0.10 and < 0.15 ng/mL ² ). Logistic regression analyses identified independent predictors of PCa and csPCa. Results: The cohort comprised 172 patients with PI‐RADS 1, 126 with PI‐RADS 2, and 175 with PI‐RADS 3 index lesions. In PI‐RADS 1–2, csPCa was detected in 2.2% (1/46) of patients with PSA < 3, whereas for PSAD < 0.10, the CDR was 5.3% (10/190). Avoided biopsy rates were 15.4% (46/298) and 63.8% (190/298), and missed csPCa rates were 3.3% (1/30) for PSA < 3% and 33.3% (10/30) for PSAD < 0.10, respectively. In PI‐RADS 1–3, csPCa was detected in 5.6% (4/76) of patients with PSA < 3 ng/mL, while for PSAD < 0.10, CDR was 8.8% (26/295). Biopsy avoidance was 16.1% (76/473) for PSA < 3 ng/mL and 62.4% (295/473) for PSAD < 0.10, with missed csPCa rates of 5.6% (4/71) and 36.6% (26/71), respectively. On multivariable analysis, higher log ₂ (PSAD) was independently associated with csPCa in both PI‐RADS 1–2 (OR, 2.171; 95% CI, 1.255–3.757; p = 0.006) and PI‐RADS 1–3 (OR, 2.727; 95% CI, 1.844–4.033; p < 0.001) cohorts. In patients with index PI‐RADS 1–3 findings, prior negative biopsy was independently associated with lower odds of csPCa (OR = 0.425; 95% CI, 0.182–0.991; p = 0.048). Non‐diagnostic MRI quality was independently associated with PCa detection in the PI‐RADS 1–2 cohort (OR, 2.367; 95% CI, 1.276–4.388; p = 0.006). Conclusions: In men with index PI‐RADS 1–3 lesions, PSA < 3 ng/mL identified a very low‐risk subgroup with low csPCa detection, and few missed csPCa. Conventional PSAD cutoffs performed worse for biopsy deferral, with a higher rate of missed csPCa. Biopsy decisions for PI‐RADS 1–3 lesions are best guided by a composite assessment that incorporates PSA, PSAD, PI‐RADS category, prior biopsy status, MRI quality, and lesion burden.

Background/Objectives: MRI is a non-invasive tool which can be used to assess baseline gonadal anatomy, including changes during puberty. Volumetric characterization offers valuable insights about the reproductive system and gonads, but annotation is cumbersome, and no AI tool is currently available. This study aimed to develop two open-source AI models to segment bilateral gonads at MRI scans in healthy subjects. Materials and Methods: This study uses a longitudinal dataset consisting of 182 MRIs from 22 healthy girls (median age 13) and 266 MRIs from 44 healthy boys (median age 13) from a single institute. MRI acquisition included T2-weighted (T2W) sequence, along with fat-saturated (FS) T2W when indicated. An expert radiologist segmented gonadal anatomy, including ovarian cysts (>3 cm). Three-dimensional nnUnet models were trained for ovary, cyst, and testicle segmentation, respectively. The ovary–cyst segmentation model was applied to an external dataset with 30 adult subjects. Model performance was evaluated on the test set using the Dice similarity coefficient for ovary (DSCOV), cyst (DSCCY), and testicles (DSCTS). Subject-level total volumes for ovaries (TOV), cysts (TCV), and testicles (TTV) were computed. Results: Ovary, cyst, and testicle segmentation models achieved DSCOV of 0.86, DSCCY of 0.69, and DSCTS of 0.90 in the in-house test set, respectively. Average mean difference with 95% confidence intervals for TOV, TCV, and TTV were 0.87 (−5.78, 7.5), −0.41 (−3.3, 2.5), and 0.19 (−1.5, 1.9) cm³, respectively. Conclusions: The developed models show promising and reliable performance in volumetric and morphologic evaluation of gonads during puberty.

Representativeness of study participants

Patient characteristics of the NCCS EBRT cohort

Patient characteristics of the combined RTOG 9202, 9413, and 9902 cohorts

Associations between MRI tumor volumes and transcriptional signatures

Multivariable analysis tables of the combined RTOG 9202, 9413, and 9902 cohorts

Gene list of a prostate cancer-specific TGF-β activity signature

Radiomic features of radiation response