Predicting inguinal lymph node metastasis in penile squamous cell carcinoma, from imaging, molecular biomarkers to multimodal AI: a narrative review

Abstract

Background and Objective

Penile squamous cell carcinoma (PSCC) is a relatively uncommon malignancy that significantly impairs patients’ quality of life. Inguinal lymph node metastasis (ILNM) is a key prognostic determinant of survival. Accurate preoperative ILNM prediction remains a major clinical challenge, emphasizing the need for better risk stratification. This review evaluates the conventional predictors and explores the potential of artificial intelligence (AI) models to enhance predictive accuracy for ILNM.

Methods

In our narrative review, the literature search was conducted in PubMed/MEDLINE, Web of Science, and Google Scholar from January 2005 to July 2025. We used the search terms: penile cancer, penile squamous cell carcinoma, inguinal lymph node metastasis, predictors, artificial intelligence, and multimodal prediction. Relevant titles and abstracts were screened; eligible full texts were reviewed.

Key Content and Findings

Conventional clinicopathological predictors and existing predictive models exhibit limited accuracy in predicting ILNM in PSCC. Current imaging techniques, such as ultrasound, computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography/computed tomography (PET/CT) provide complementary information, but each modality has limitations. Molecular and genomic biomarkers offer biological insights, but validation remains inconsistent. Recent AI approaches that integrate diverse data types have demonstrated superior predictive performance compared to unimodal models. Multimodal AI-based techniques have the potential to improve personalized risk stratification and inform management strategies. However, clinical adoption of AI-based frameworks in penile cancer is limited by major challenges, including data scarcity, heterogeneity, and lack of standardization. Addressing these issues through prospective multicenter cohorts with external validation will facilitate AI integration in clinical practice.

Conclusions

ILNM is a critical prognostic factor and current predictors are suboptimal. Integrating clinicopathological, molecular, and imaging features through AI-based multimodal frameworks may enhance ILNM prediction and guide surgical decision-making. However, AI applications in penile cancer are in the early stages and require validation through large, multicenter trials.

Introduction

Penile squamous cell carcinoma (PSCC) is a rare malignancy, with significant variation in incidence rates across regions (1). In the USA and Europe, incidence rates range from 0.5 to 0.94 per 100,000 men; however, developing nations (e.g., Brazil, India, and some parts of Africa) bear a disproportionately high burden, with rising trends (2). Among prognostic factors, inguinal lymph node metastasis (ILNM) is a well-established determinant of survival outcomes (3). Five-year cancer-specific survival rates decline from over 85% in node-negative disease to less than 20% in patients with advanced nodal involvement (4). Early surgical management of inguinal lymph nodes significantly improves survival outcomes (84% compared with 35% for delayed intervention) (5). Inguinal lymph node dissection (ILND) has long been the cornerstone of treatment for PSCC, providing accurate staging and reducing mortality. However, it is associated with high complication rates, reaching up to 70% (6). Emerging minimally invasive techniques, such as videoendoscopic ILND (VEIL), robotic-assisted VEIL (RA-VEIL), robotic-assisted ILND (RA-ILND), and single-port approaches, show promise in reducing morbidity, enhancing recovery, and achieving comparable oncological outcomes. These advancements offer superior cosmetic results and fewer complications, yet further validation is needed, particularly in complex cases involving pelvic lymphadenectomy (7-9).

Current predictive methods for ILNM mostly rely on clinicopathological parameters, nomograms, and conventional imaging (10). Imaging modalities [ultrasound (US), magnetic resonance imaging (MRI), and positron emission tomography/computed tomography (PET/CT)] assist in evaluating ILNM in PSCC, but they show limited sensitivity in detecting occult metastases and cannot reliably distinguish malignant from reactive lymph nodes. This often leads to staging inaccuracies (11). To address critical gaps in preoperative nodal assessment, emerging evidence highlights the importance of advanced multimodal risk stratification, incorporating molecular biomarkers and hybrid imaging (PET-MRI/CT) (12). Artificial intelligence (AI) has improved predictive accuracy in common urologic cancers; however, its application to ILNM prediction in penile cancer remains in the early stages, primarily due to the rarity of the disease. This review consolidates current evidence on clinicopathological, imaging, and molecular predictors of ILNM, and explores the role of multimodal AI frameworks in penile cancer. It also critically examines methodological limitations arising from scarce datasets and advocates data-augmentation strategies to address these challenges. We present this article in accordance with the Narrative Review reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-742/rc).

Methods

This narrative review was conducted using a structured search strategy across multiple databases, including PubMed/MEDLINE, Web of Science, and Google Scholar, covering the period from January 2005 to July 2025. Relevant studies were identified by searching for key terms such as “penile cancer”, “penile squamous cell carcinoma”, “inguinal lymph node metastasis (ILNM)”, “predictors”, “imaging biomarkers”, “molecular biomarkers”, “genomic biomarkers”, “radiomics”, “radiogenomics”, “multimodal prediction”, and “artificial intelligence”. The search was limited to articles published in English. Eligible studies were selected based on their relevance to ILNM prediction in PSCC, focusing on clinicopathological, imaging, molecular, and AI-based predictors. Original research, meta-analyses, and high-quality reviews with methodological rigor, adequate sample size, and clinical relevance were prioritized to address conflicting results (Table 1).

Table 1. The search strategy summary.

Items	Specification
Date of search	20 July 2025 to 12 August 2025
Databases and other sources searched	PubMed/MEDLINE, Web of Science, and Google Scholar
Search terms used	Key terms: “penile squamous cell carcinoma”, “penile cancer”, “inguinal lymph node metastasis (ILNM)”, “predictors”, “imaging biomarkers”, “molecular biomarkers”, “genomic biomarkers”, “radiomics”, “radiogenomics”, “multimodal prediction”, and “artificial intelligence”. The Boolean operator ‘AND’ was used to combine core concepts, and ‘OR’ was used for synonyms within each concept
Timeframe	Studies published from 01 January 2005 to 01 July 2025
Inclusion and exclusion criteria	Research articles, systematic reviews, and meta-analyses published in English between 01 January 2005 and 01 July 2025, focusing on ILNM prediction in PSCC were included
	Studies focusing on non-squamous penile tumors, case reports, published in languages other than English and studies that did not provide adequate data on ILNM prediction were excluded
Selection process	Study selection was conducted by M.A.; all authors reviewed the process and reached consensus
Any additional considerations	Ethical approval was not required due to the narrative review design
ILNM, inguinal lymph node metastasis; PSCC, penile squamous cell carcinoma.

Conventional prediction approaches

Clinicopathological predictors

Palpable inguinal lymph nodes remain the most crucial clinical predictor of ILNM. Initial assessment should include evaluating the site and size of the primary lesion, careful palpation of both groins, and documentation of the number, laterality, and characteristics of any suspicious lymph nodes (5). Clinically node-positive (cN+) groins carry a high probability (45–80%) of metastatic disease (13), whereas clinically node-negative (cN0) groins appear disease-free but still harbor occult metastases in 20–25% of cases (3). Furthermore, several studies have reported a significant association between age at diagnosis and ILNM (6,14). Imaging modalities, including ultrasonography and MRI, serve as adjuncts in cases with difficult examination or staging uncertainty, though routine imaging is not recommended for cN0 patients.

Several histopathological factors have been identified as independent predictors of ILNM, such as lymphovascular invasion (LVI), poor histological T grade, higher pathological T stage, cN+, and perineural invasion (PNI) (15,16). A recent meta-analysis conducted by Hu et al., containing 42 studies and 4,802 patients, identified several significant clinicopathological predictors of ILNM (P<0.05) (17). Among all, LVI has been most consistently reported as the strongest independent predictor of ILNM in PSCC (18). A previous study identified that, despite being cN0 and tumor size, both LVI+ and PNI+ are important predictors of ILNM (19). Another study indicated that patients with corpora cavernosa invasion had a greater metastatic rate compared to those with corpora spongiosa invasion, as per the eighth edition tumor stage (20). The predictive value of various clinical and pathological parameters is comprehensively summarized in Table 2.

Table 2. Clinicopathological predictors of ILNM and nomograms in PSCC.

Study	Key predictors of LNM	OR (95% CI)	P value (multivariate)	Nomogram (AUC)/c-index	Interpretation
	Age (years)			0.776	Multivariate analysis showed that primary tumor site, higher nuclear grade, tumor size >3 cm, and higher T stage were identified as being significantly associated with lymph node involvement. However younger age (<50 years) significantly reduces the odds of LNM compared to older age groups (50–69 and ≥70 years)
Zhang et al. (2021) (16)	<50 vs. 50–69	0.534 (0.331–0.863)	0.01
	<50 vs. ≥70	0.357 (0.215–0.594)	<0.0001
	Tumor site
	Glans penis	1.122 (0.933–2.563)	0.069
	Body of the penis	1.456 (1.013–1.565)	0.034
	Overlapping lesions	1.613 (1.035–1.781)	0.023
	pT stage
	T2 vs. T1	3.717 (2.306–5.991)	<0.0001
	T3–T4 vs. T1	6.173 (3.677–10.362)	<0.0001
	Higher grade
	G2 vs. G1	2.84 (1.679–4.805)	<0.0001
	G3–4 vs. G1	5.629 (3.201–9.900)	<0.0001
	Tumor size >3 cm	1.461 (1.017–2.098)	<0.0001
Peak et al. (2019) (15)	Higher grade			0.880	Study has demonstrated on multivariate analysis, LVI, clinical nodes positive (cN+), and higher nuclear grade are significant predictors of ILNM
	G2 vs. G1	2.58 (1.39–4.79)	0.002
	G3/4 vs. G1	3.27 (1.70–6.29)	0.002
	LVI	2.49 (1.61–3.84)	<0.0001
	cN+		<0.0001
	N1 vs. N0	20.0 (11.4–35.7)
	N2 vs. N0	27.8 (14.1–55.6)
	N3 vs. N0	49.2 (14.8–162.8)
Fankhauser et al. (2022) (19)	LVI			NR	Either LVI+ or PNI+ are the significant predictive factors for ILNM despite being cN0 and tumor size
	LVI+ vs. LVI−	3.8 (1.4–9.0)	<0.01
	PNI+ vs. PNI−	6.6 (2.2–18.0)	<0.01
	LVI+ or PNI+ vs. LVI− or PNI	3.9 (1.6–9.0)	<0.01
	Tumor size	1.01 (0.99–1.04)	0.17
Qu et al. (2018) (21)	Age (≥60 years)	0.68 (0.52–0.88)	0.0035	NR	Study confirmed and externally validated that the age, stage, and grade of the primary tumor are significant predictors of lymph node metastasis
	pT (≥pT1b vs. <pT1b)	3.32 (1.38–8.01)	0.0075
	Tumor grade
	Grade 2 (vs. 1)	2.98 (1.26–7.62)	0.023
	Grade 3 (vs. 1)	3.97 (1.32–11.9)	0.014
Wu et al. (2021) (10)	SCC-Ag	1.090 (1.035–1.148)	0.001	0.817	Proposed a clinicopathologic and laboratory-based nomogram that incorporates PLR, SCC-Ag, LVI, and pT stage, can accurately predict ILNM/ENE in patients with penile cancer
	PLR	1.012 (1.006–1.019)	<0.001
	pT stage
	pT2	6.522 (1.716–24.791)	0.006
	pT3	8.077 (2.322–28.10)	0.001
	pT4	23.258 (2.43–222.56)	0.006
	LVI	3.205 (1.227–8.371)	0.017
Zhou et al. (2020) (22)	Grade (G3–4 vs. G1–2)	6.467 (1.241–33.684)	0.027	0.948	Tumor grade, LVI, and short diameter of the largest ILN were independent risk factors on multivariate analysis. Nomograms can efficiently predict the ILNM in penile cancer patients with cN+
	LVI (present vs. absent)	5.162 (1.056–25.243)	0.043
	SD of the largest node	1.349 (1.133–1.606)	0.001
Unadkat et al. (2021) (23)	pT stage	0.97 (0.35–2.75)	0.95	NR	On multivariable analysis, only higher tumor grade and LVI were independently associated with occult ILNM, whereas pT stage was not
	pT1b	1.10 (0.51–2.55)	0.81
	pT3/T4	1.21 (0.53–2.93)	0.66
	Tumor grade
	G2 (vs. 1)	2.16 (1.16–4.32)	0.02
	G3–4 (vs. 1)	2.81 (1.40–5.94)	0.005
	LVI	3.12 (1.85–5.27)	<0.001
Nascimento et al. (2020) (24)	LVI	7.224 (0.831–22.730)	0.029	NR	Presence of LVI and absence of koilocytosis were found to be predictive factors for LNM on multivariate analysis
	Absent koilocytosis	0.088 (2.628–50.718)	0.001
Hu et al. (2019) (17)	Histopathological subtype			NR	Meta-analysis identified valuable predictors of ILNM in penile cancer patients, including higher stage, LVI, PNI, corpora cavernosa invasion, urethral invasion, larger tumor size, higher grade, vertical growth pattern, and high- and intermediate-risk histopathological subtype
	High-risk	14.63 (6.40–33.42)	<0.001
	Intermediate-risk	3.37 (1.97–5.74)	<0.001
	Vertical growth pattern	1.97 (1.13–3.43)	0.020
	Higher T stage (AJCC)	3.66 (2.47–5.42)	<0.001
	Higher T stage (UICC)	2.43 (1.80–3.26)	<0.001
	Higher tumor grade	3.37 (2.38–4.78)	<0.001
	Depth of invasion	2.58 (1.42–4.64)	0.002
	LVI	2.88 (2.20–3.75)	<0.001
	Corpora cavernosa invasion	2.22 (1.63–3.04)	<0.001
	Corpus spongiosum invasion	1.73 (1.22–2.46)	0.002
	Urethral invasion	1.81 (1.07–3.05)	0.030
	Nerve invasion	2.84 (1.99–4.04)	<0.001
Zhang et al. (2024) (25)	Lymph node diameter	1.366	0.001	0.81	Developed a multifactorial predictive model for ILNM that demonstrated good predictive performance, with an area under the curve of 0.8
	T stage	1.134	NR
	NLR	1.034	0.002

AJCC, American Joint Committee on Cancer; AUC, area under the curve; CI, confidence interval; cN, clinical node status; ENE, extranodal extension; ILN, inguinal lymph node; ILNM, inguinal lymph node metastasis; LNM, lymph node metastasis; LVI, lymphovascular invasion; NLR, neutrophil-to-lymphocyte ratio; NR, not reported; OR, odds ratio; PLR, platelet-to-lymphocyte ratio; PNI, perineural invasion; PSCC, penile squamous cell carcinoma; pT, pathological T stage; SCC-Ag, squamous cell carcinoma antigen; SD, short diameter; UICC, Union for International Cancer Control.

Despite the wide range of clinicopathological predictors reported, their relative predictive performance and clinical utility vary substantially. Among these factors, LVI and pathological T stage consistently demonstrate the strongest and most reproducible associations with ILNM across multiple cohorts and meta-analyses. In contrast, variables such as age, tumor size, and primary tumor location show less consistent effects and are more susceptible to confounding. Importantly, most clinicopathological predictors rely on postoperative pathological assessment, which limits their value for preoperative decision-making and risk stratification.

Tumor-node-metastasis (TNM) staging and current nomograms for risk stratification

The TNM staging system continues to serve as a basic tool for staging, prognosis, and treatment planning in penile cancer. The American Joint Committee on Cancer (AJCC) released its 8th edition in 2017. The AJCC 8th edition staging system for penile cancer improves upon previous versions by incorporating extranodal extension (ENE), subdividing T1 into T1a and T1b on the basis of LVI/PNI, and distinguishing invasion of the corpus spongiosum as pT2 from corpus cavernosum invasion as pT3, regardless of urethral involvement (5). The AJCC 8th edition staging system effectively stratifies disease-specific survival (DSS); nevertheless, a multicenter investigation showed that a modified model offered improved predictive accuracy and more distinct stage discrimination, highlighting areas for refinement (26). Additionally, it primarily describes anatomical tumor extent but excludes molecular markers and advanced imaging, which limits its preoperative predictive accuracy (27).

In penile cancer, nomograms are considered useful tools for personalized metastatic risk estimation (6). Nomograms incorporate readily accessible clinicopathological factors, such as age at diagnosis, T grade, T size, and primary tumor site, which assist in medical decisions and patient counseling with moderate to good accuracy (16). Nomograms mainly rely on postoperative pathology, and the exclusion of key determinants such as imaging and molecular biomarkers [human papillomavirus (HPV) status] limits their predictive accuracy (28). Most nomograms are based on single-institution cohorts, lack external validation, and omit key variables, which significantly restrict their generalizability and clinical utility (29). Several key nomograms with their predictive performance are described in Table 2.

Dynamic sentinel node biopsy (DSNB) and surgical staging

DSNB has emerged as a minimally invasive staging method that leverages the stepwise pattern of metastatic spread in PSCC (30), enabling precise nodal staging in cN0 patients (31). Among all surgical staging techniques, DSNB offers the lowest morbidity and the highest diagnostic accuracy (sensitivity 92–96%, false-negative rate 4–8%, complication rate 6–14%) (32). The current European Association of Urology (EAU) guidelines advocate DSNB for cN0 patients and patients with intermediate (pT1a G2) to high-risk tumors (≥ pT1b) (5). DSNB should be conducted exclusively in specialized centers with expertise in lymphatic mapping and penile cancer to maximize accuracy and reduce false-negative rates. However, concerns regarding cost-benefit remain (32). Moreover, recent developments in intraoperative procedures and imaging tracers hold promise to lower false-negative rates and increase safety (33). In sentinel node positive cases, ILND is indicated.

ILND continues to be the standard staging and therapeutic approach for cN+ patients and provides important prognostic information, such as ENE and nodal burden (34). Despite its potential advantage in tumor grading and lowering mortality risk, it remains a controversial procedure (35). ILND is associated with high complication rates, including lymphedema, infection, delayed healing, and skin necrosis, which have a significant impact on long-term quality of life (35). Consequently, the accurate preoperative identification of patients who will benefit from ILND is imperative to avoid unnecessary morbidity. Several approaches have been proposed to mitigate these adverse outcomes. Modified inguinal lymphadenectomy (mILND) reduces lymphatic complications (10–36%) compared to radical dissection while maintaining oncologic efficacy in high-risk cN0 patients with saphenous vein preservation (36). Video-endoscopic inguinal lymphadenectomy (VEIL) demonstrates lower morbidity, better lymphatic preservation, and shorter hospital stay compared with open surgery. VEIL provides comparable short-term oncologic outcomes (37). However, long-term oncological safety necessitates further validation.

Contemporary imaging modalities

Ultrasonography

US serves as a primary imaging tool for lymph node examination, offers real-time and cost-effective evaluation, yet its accuracy is limited when used alone (4). US-guided core biopsy or fine needle aspiration cytology (FNAC) in cN+ cases to confirm nodal metastasis before proceeding to surgery (ILND) is useful (5). Furthermore, US is a practical alternative to MRI for evaluating primary tumor invasion in the corpus spongiosum or cavernosum. When combined with techniques such as DSNB, its accuracy improves, enhancing staging in both cN0 and cN+ nodes and providing additional surgical and logistical advantages (38). Contrast-enhanced ultrasonography (CEUS) is an emerging technique that has been used in various cancers, but has not been applied in penile cancer (11).

MRI

MRI has become increasingly significant for local staging in penile cancer following the 8th AJCC-TNM revisions. The new EAU guidelines recommend MRI to evaluate penile anatomy and tumor extension, particularly to detect corpora cavernosa invasion (39). In a small cohort, Lucchesi et al. detected ILNM in 13/15 cases (86.7%) with MRI versus 7/15 (46.7%) by physical examination (40). MRI outperforms palpation in nodal staging, especially when physical examination is suboptimal, such as in obese patients (40). However, palpation remains the preferred approach for decision-making. Accurate detection of occult metastases and distinction between reactive and malignant palpable nodes remains challenging. Although the use of novel contrast agents such as ultra-small superparamagnetic iron oxide particles (USPIO) (41) and PET-MRI holds potential to address these limitations (42).

The apparent diffusion coefficient (ADC) derived from diffusion-weighted imaging (DWI)-MRI is an emerging biomarker; lower values indicate high cellular density, a hallmark of metastatic potential. A recent study reported that reduced ADC values in primary tumors predict adverse histology and nodal metastasis (87.2% sensitivity, 80.9% specificity), surpassing conventional imaging (43). Dynamic contrast-enhanced (DCE)-MRI assesses tumor vascularity by quantifying perfusion parameters (K_trans) that reflect angiogenesis and metastatic potential. However, its clinical utility is limited by scarce evidence and lack of standardized protocols (44).

PET/CT

Fluorine-18 fluorodeoxyglucose (¹⁸F-FDG) PET/CT combines metabolic information with anatomical imaging and provides enhanced diagnostic accuracy. Primary penile tumors typically show high FDG uptake, yet not routinely recommended for local staging owing to its limited spatial resolution (12). ¹⁸F-FDG PET/CT has demonstrated a sensitivity of 96% in cN+ patients and only 57% in cN0 cases (45). In penile cancer with confirmed inguinal metastases, ¹⁸F-FDG PET/CT showed high accuracy for pelvic nodal staging (sensitivity: 91%, specificity: 100%, accuracy: 96%) (46). Although FDG-PET/CT outperforms conventional imaging in advanced disease (47), it is not recommended for cN0 cases owing to limited spatial resolution for detecting small metastases. The maximum standardized uptake value (SUVmax) is another important quantitative marker that reflects tumor metabolic activity and provides critical prognostic insights; higher values indicate aggressive malignancy (12). However, PET/CT is prone to false positives from inflammatory uptake, so SUV metrics should be interpreted cautiously and corroborated with additional diagnostic techniques (48).

Multimodal imaging integration

Conventional imaging techniques provide useful information but remain limited in sensitivity and specificity for detecting lymph node metastases when applied in isolation. For instance, to detect small nodal metastases, MRI shows suboptimal accuracy in staging, with persistent false-negative rates (39). ¹⁸F-FDG PET/CT provides valuable metabolic insight, yet it is hindered by limited spatial resolution and potential false positives from inflammatory uptake (35). Multimodal imaging techniques that incorporate anatomical (CT/MRI), functional (DWI/DCE-MRI), and metabolic (PET/CT) data show promise for accurate detection of lymph node metastasis. In a group of 26 penile cancer patients with palpable inguinal nodes, the combination of sentinel lymph node biopsy (SLNB), ultrasound, and single photon emission computed tomography/CT (SPECT/CT) detected five extra metastatic groins that were missed by SLNB alone (49). Consequently, integrating complementary imaging modalities may augment diagnostic precision and improve surgical decision-making. Nevertheless, large prospective studies are needed for further validation.

Although advanced imaging techniques offer incremental improvements over conventional modalities, their reproducibility and clinical adoption remain constrained. Variability in acquisition protocols, reader expertise, and interpretive criteria contributes to inconsistent performance across institutions. Furthermore, while functional and metabolic imaging parameters such as ADC values and SUV metrics show promise, their routine use is limited by lack of standardized thresholds and prospective validation. As a result, imaging currently serves primarily as a complementary rather than definitive tool for preoperative nodal risk stratification.

Molecular biomarkers & genomic insights

Immunohistochemistry (IHC) biomarkers

HPV plays a major role in PSCC, being implicated in 23–50% of overall penile cancer cases (50). The most common subtypes are HPV16 and HPV18 (51). HPV-positive tumors are commonly identified by p16INK4a overexpression on IHC, which serves as a surrogate for HPV positivity (52). Currently, the WHO 2022 classification divides PSCC into HPV-associated or HPV-independent (53). HPV-positive tumors demonstrate lower ILNM rates (15–25% vs. 35–50% in HPV-negative cases) and better overall survival (OS) (54). According to the updated American Society of Clinical Oncology (ASCO)-EAU guidelines, HPV testing in the initial diagnostic workup of PSCC carries a level 2a evidence grade, though additional studies are required to justify its universal implementation (5).

An immune checkpoint marker known as programmed cell death ligand 1 (PD-L1) has shown a substantial correlation between nodal involvement and upregulation (55). Hu et al. identified PD-L1 as an independent predictor of ILNM on multivariate analysis [odds ratio (OR) 5.16, P=0.02] (56). p53 overexpression reflects underlying TP53 alterations and has been associated with nodal metastasis in penile cancer (57). Prapiska et al. demonstrated that higher p53 expression was correlated with markedly reduced 3-year survival up to 18% (58). However, the variability in IHC assessment limits its reliability as a standalone biomarker. Ki-67 is a nuclear proliferation marker that has been investigated as a predictor of ILNM in PSCC (59). Studies have shown that Ki-67 expression correlates with nodal involvement in univariate analyses, but multivariate analyses consistently fail to confirm its independent prognostic value (60). Key molecular biomarkers of ILNM and their clinical relevance in PSCC are summarized in Table 3.

Table 3. Predictive molecular biomarkers of ILNM in PSCC.

Biomarker	Assessment method	Association with ILNM	Clinical relevance
HPV (DNA/RNA)	ISH	HPV-related tumors correlated with lower rates of metastatic disease (P=0.007) (61). Nonsignificant association between HPV positivity and nodal disease (59)	Presence of HPV which is found in 30–50% of cases lead to better prognosis and treatment strategies
p16 status	IHC	p16INK4a shows paradoxical association with occult LNM, however lacked statistical significance (62)	Better prognosis in patients with elevated expression (63). leveraging p16INK4a detection via IHC is a cost-effective surrogate marker for HPV (64)
Ki-67	IHC	Significant correlation between the Ki-67 expression and lymph node metastasis in PSCC (P=0.045) (65), but independent associations are inconsistent across studies (59)	Higher expression is accompanied by a worse prognosis
TP53	IHC	p53 is a predictor of LNM, irrespective of p16 status (57). p53 positivity is associated with LNM on uni- and multivariable analyses (OR: 6.01–22.431, P<0.01) (59)	Higher expression is accompanied by a worse prognosis
PD-L1	IHC	PD-L1 overexpression has significant association with LNM (56), (OR 2.55, 95% CI 1.40–4.64; P=0.002) (17)	High expression correlates with regional LNM and reduced CSS (66), and worse prognosis. Potential therapeutic target in penile carcinoma (NCT04224740) (67)
SCC-Ag	Serological	Elevated preoperative SCC-Ag has significant association with LNM on univariate (68), and multivariate analysis (10)	Utility for detecting LNM remains controversial and insufficiently reliable for routine clinical use
CRP	Serological	Raised CRP levels are significantly associated with LNM on univariate analysis (68)	Elevated CRP levels are linked to poorer prognosis and may serve as a prognostic indicator in penile cancer (69)
NLR	Serological	Significantly associated with LNM on univariate (70), and multivariate analyses (49), (OR 4.22, 95% CI: 1.36–13.09; P=0.010) significant predictor of LNM (17)	NLR is an independent predictor of ILNM in penile cancer (56)
Micro RNAs	qRT-PCR, microarray, next-generation sequencing	Overexpression of miR-421 have significant association with ILNM (71). Upregulation of miR-223-3p, miR-107, and miR-21-5p significantly linked with LNM (72)	Dysregulation of miRNAs reveals mechanisms involved in tumor progression and serves as a potential prognostic biomarker, predictive of metastatic disease and shorter OS

CI, confidence interval; CRP, C-reactive protein; CSS, cancer-specific survival; HPV, human papillomavirus; IHC, immunohistochemistry; ILNM, inguinal lymph node metastasis; ISH, in situ hybridization; LNM, lymph node metastasis; miRNA, microRNA; NLR, neutrophil-to-lymphocyte ratio; OR, odds ratio; OS, overall survival; PD-L1, programmed death-ligand 1; PSCC, penile squamous cell carcinoma; qRT-PCR, quantitative reverse transcription polymerase chain reaction; SCC-Ag, squamous cell carcinoma antigen.

Serological and other systemic inflammatory biomarkers

Serum squamous cell carcinoma antigen (SCC-Ag) levels often rise in patients with nodal metastases prior to clinical detection (73). Elevated preoperative SCC-Ag (>1.4 ng/mL) and C-reactive protein (CRP) (≥4.5 mg/L) were both significantly linked with lymph node metastasis laterality and extra nodal extension (EXE) (P<0.001). Moreover, their combined positivity independently predicted adverse outcomes [hazard ratio (HR) 3.39, P=0.033] (68). However, SCC-Ag remains a poor predictor and needs validation in larger cohorts (59). High preoperative serum CRP levels (≥20 mg/L) have been associated with ILNM and poorer survival outcomes (74). A recent meta-analysis of 989 patients confirmed that elevated CRP was linked to increased risk of lymph node metastasis (LNM) (risk ratio 2.27, P<0.01) and poorer prognosis (69).

Among systemic inflammatory indices, the neutrophil-to-lymphocyte ratio (NLR) has been extensively studied as an independent predictor of ILNM in penile cancer (56,75). Azizi et al. reported NLR ≥3 was associated with an increased risk of ILNM (OR 3.75, P=0.014) on univariable analysis (76), though independent predictive value remains uncertain. Platelet-to-lymphocyte ratio (PLR) is another predictor that has been identified as a significant independent predictive factor for ILNM (77). PLR demonstrated significant association with ENE in penile cancer (10). The systemic immune-inflammation index (SII), a combination of neutrophil, lymphocyte, and platelet counts, has shown promise as a comprehensive inflammatory predictive biomarker of ILNM in PSCC. Song et al. found that higher preoperative SII levels (cut-off 636.99) were significantly associated with total penectomy and markedly worse survival (median OS 10.5 vs. 128 months) (78).

Genomic alterations

The molecular landscape and tumor microenvironment in PSCC have been recently elucidated by genomic profiling, especially next-generation sequencing (79). The genes most frequently mutated in penile cancer and associated with tumor aggressiveness and metastatic risk include TP53, FGFR3, CDKN2A, PIK3CA, NOTCH1, MYC, HRAS, and EGFR (80). TP53 mutations, found in approximately 38–45% of cases, are strongly associated with advanced nodal disease and poor prognosis, especially in HPV-negative tumors (81).

In the last few years, the use of microRNAs as biomarkers has gained attention in the detection and screening of various malignancies. A recent study reported that overexpression of miR-223-3p, miR-107, and miR-21-5p correlated with poor prognosis; miR-223-3p upregulation was also associated with nodal involvement (72). Ayoubian et al. identified that lower expression of miR-137 and miR-328-3p is indicative of metastatic disease (82). The majority of transcriptomic and genomic data comes from small, retrospective cohorts, which emphasizes the need for multicenter validation to determine their clinical utility.

While molecular and genomic biomarkers provide important biological insights into metastatic potential, their translation into clinical practice remains limited. Many proposed biomarkers demonstrate associations with nodal involvement in single-center or retrospective studies, but few have undergone external validation or prospective testing. Variability in assay techniques, cut off definitions, and patient populations further limits reproducibility. Consequently, molecular biomarkers currently function as adjunctive risk indicators rather than standalone predictors for clinical decision-making.

AI and multimodal integration

Radiomics and radiogenomics

Radiomics is an emerging field that leverages AI to extract high-dimensional imaging features (e.g., tumor heterogeneity, texture) from medical imaging (MRI and PET/CT) that is beyond the limits of human visual perception (83,84). In a 38-patient cohort, expert-applied Graafland CT criteria predicted high-risk nodal disease with 86.8% accuracy (89.5% inter-reader concordance). A two-feature CT radiomics model reached area under the curve (AUC) 0.921 (five-fold cross-validation), indicating added value for standardized risk stratification (85). Radiogenomics is a multidisciplinary approach that combines imaging-derived features with molecular and genomic data to enhance cancer characterization and risk stratification (83). Unlike other major urological malignancies, radiomics and radiogenomics in penile cancer remain exploratory, constrained by disease rarity and limited annotated multimodal datasets.

AI diagnosis in penile cancer

AI has advanced urologic oncology by supporting predictive modeling for diagnosis, prognostication, and individualized treatment planning (86). However, AI application to predict nodal metastasis in penile cancer remains markedly limited due to its rarity. A machine learning (ML) study compared five algorithms and found the eXtreme Gradient Boosting (XGB) model achieved the highest accuracy for ILNM prediction (AUC 0.853) (87). Among all deep learning (DL) techniques, convolutional neural networks (CNNs) are the most frequently used architectures for medical image analysis (CT, MRI, pathology slides) (88).

The CNN architectures automatically extract hierarchical features and preserve spatial relationships between tumor and normal tissue, which makes them highly effective for large-scale cancer detection and semantic segmentation tasks. An initial trial of CNN trained on 136 open-access penile lesion images demonstrated high diagnostic accuracy in distinguishing benign from malignant lesions (AUC 0.94, sensitivity 82%, specificity 87%), but only moderate accuracy for precancerous lesions (AUC 0.74) (89). Similarly, another study using U-Net and Inception-ResNet v2 reached an accuracy of 0.944 when classifying five penile conditions from clinical pictures, suggesting potential for mobile health apps (90).

AI applications in digital histopathology show promising potential to enhance diagnostic accuracy in penile cancer. A study by R et al. reported that optimized neural architecture search (NAS) models outperformed conventional approaches and achieve F1-scores of 89.5% (40×) and 88.5% (100×) (91). Two studies by Lauande et al. developed DenseNet-based methods for penile cancer diagnosis on histopathological images. One achieved an F1-Score of 97.39% and a sensitivity of 98.33% in classifying normal tissue and SCC (92). While the other study, used a DenseNet enhanced with Transformer and Mobile Inverted Bottleneck Convolution (MBConv) blocks, pretrained on other cancer datasets, achieved an F1-score of 95.8% (93). Collectively, these results highlight the potential of AI for penile cancer diagnostics, though most studies remain limited due to small datasets and lack of external validation.

Despite these encouraging early results, it is important to emphasize that most AI applications in penile cancer remain at a proof-of-concept stage. Existing studies are predominantly retrospective, single-center, and based on limited sample sizes, which increases the risk of overfitting and limits generalizability. At present, these AI tools should be viewed as research instruments rather than clinically deployable decision support systems.

Multimodal AI frameworks

Cancer management generates extensive multimodal data derived from imaging (CT, MRI, PET), histopathology, genomic profiling, and clinical documentation (94). Multimodal fusion models are advanced AI algorithms designed to integrate the diverse data sources and typically outperform single-modality approaches in predictive accuracy (95). Task-specific DL models are employed for multimodal data processing. The CNN architecture analyzes imaging data (CT/MRI/WSI) via pixel matrices, while sequential models such as recurrent neural networks (RNNs) or transformers process and convert clinical text via word embeddings (96). Each modality requires tailored preprocessing before multimodal fusion. Based on integration timing, multimodal fusion techniques can be categorized as early, intermediate, and late fusion techniques (97). Recent advances in multimodal DL have significantly enhanced predictive accuracy across various cancers by integrating heterogeneous data (98,99). To date, multimodal AI frameworks have not been directly applied to ILNM prediction in penile cancer. However, their proven success in major urological cancers (100) suggests strong potential for improving ILNM risk stratification and personalized management. Figure 1 depicts a conceptual multimodal AI workflow for ILNM prediction in PSCC.

Figure 1.

Multimodal AI-framework for preoperative prediction of ILNM in PSCC. Radiology and H&E WSI feed a CNN/MIL image encoder, while IHC scores, genomics, and clinical variables feed an MLP structured-data encoder. The unimodal outputs undergo multimodal integration (late fusion via concatenation) to produce an ILNM probability that guides management strategies. mILND is an acceptable staging alternative to DSNB. Figure 1 was created with BioRender.com. *, DSNB should be performed in experienced centers. AI, artificial intelligence; cN0/cN+, clinically node-negative/positive; CNN, convolutional neural network; DSNB, dynamic sentinel node biopsy; IHC, immunohistochemistry; ILNM, inguinal lymph-node metastasis; MIL, multiple-instance learning; mILND, modified inguinal lymph-node dissection; MLP, multilayer perceptron; PSCC, penile squamous cell carcinoma; WSI, whole-slide image.

Although multimodal AI frameworks have demonstrated clear advantages in more prevalent urological malignancies, their application to penile cancer remains theoretical. No multimodal AI model for ILNM prediction in penile cancer has yet undergone prospective validation or clinical implementation. Therefore, current evidence supports their potential rather than their readiness for routine clinical use. The relative clinical maturity and applicability of these emerging approaches are summarized in Table 4.

Table 4. AI tools for penile cancer diagnosis and ILNM prediction.

Studies	AI tool/architecture	Modality	Purpose	Clinical applicability & limitations
Fazili et al. (2025) (85)	Radiomics (CT)	Imaging	Risk stratification in node-positive penile cancer	Not yet validated for clinical use. Needs external validation
Ding et al. (2023) (87)	ML (XGB model)	Imaging + clinical data	Predicting early-stage ILNM risk in PSCC	Early-stage potential. Small external validation data set (16.2%). Requires multicenter collaboration for validation
Liu et al. (2024) (89)	Deep learning (CNN)	Imaging (photographs)	Detection of penile cancer from photographs	Exploratory; research is needed to refine and validate the AI software with real-life data
Allan-Blitz et al. (2024) (90)	Mobile AI app (U-Net and Inception-ResNet v2)	Imaging + clinical data	Diagnosis of penile diseases using AI	Early-stage development; promising for mobile use. The model was based on clinical images rather than histopathology slides. Further evaluation on larger prospective data sets is warranted
R et al. (2025) (91)	CNN architectures	Digital histopathology	Diagnosis of penile cancer using histopathological images	Potential for clinical use. Requires further external validation
Lauande et al. (2022) (92)	DenseNet + Transfer Learning	Digital histopathology	Classifying penile cancer from histopathological images	Needs larger datasets for external validation. Currently limited by small sample size and lack of external validation
Lauande et al. (2024) (93)	DenseNet + Transformer + MBConv	Histopathology	Histopathological classification of penile cancer	limited dataset, pre-trained on other cancers. Promising, still in the early stage with unproven clinical application

AI, artificial intelligence; CNN, convolutional neural network; CT, computed tomography; DenseNet, densely connected convolutional network; ILNM, inguinal lymph node metastasis; MBConv, mobile inverted bottleneck convolution; ML, machine learning; PSCC, penile squamous cell carcinoma; ResNet, residual networks; XGB, extreme gradient boosting.

Clinical translation challenges and strategies

Clinical translation of multimodal AI models in penile cancer is primarily challenged by scarce and heterogeneous datasets, which are inherent to this rare malignancy. Additionally, non-standardized protocols across institutions further constrain the integration of multi-source data into a unified analytical framework.

Data scarcity and heterogeneity

Penile cancer accounts for <1% of male malignancies, resulting in datasets that are too small for developing robust AI models (86). Unlike common cancers with large datasets (101), penile cancer research faces data scarcity and high interobserver variability in clinicopathological predictors of ILNM. These challenges further complicate model development (17). To overcome these challenges, multicenter collaborations are imperative to aggregate heterogeneous datasets comprising imaging, histopathological, and clinical variables (4). New evidence shows that data-efficient methods like generative adversarial networks (GANs) can augment limited datasets by generating realistic synthetic histopathological and imaging data, thereby improving model performance in related urologic malignancies (102). Furthermore, transfer learning, which leverages pre-trained models on larger datasets from related domains, offers a powerful strategy to enhance predictive accuracy with limited datasets (103).

Standardization

The clinical adoption of AI is hampered by non-standardized imaging protocols and inconsistent histopathology reporting among institutions (104). Potential imaging parameters (ADC, novel contrast agents) require standardized acquisition and post-processing to achieve reproducibility and reliability. Community initiatives such as image biomarker standardization initiative (IBSI) have established consensus guidelines on image processing workflows, feature definitions, and reporting standards to improve comparability across the centers. Adopting common data models such as Observational Medical Outcomes Partnership (OMOP) enhances interoperability and facilitates multicenter AI validation. Furthermore, federated learning (FL) enables privacy-preserving, cross-institutional model development without centralizing sensitive data, improving robustness and generalizability (105,106).

Ethical considerations

AI holds transformative potential for urological oncology; however, it also raises substantive ethical concerns, including algorithmic bias, transparency, and data confidentiality (107). As AI applications in urology contain sensitive data, it demands strict privacy standards that must be implemented by AI developers and healthcare professionals (108). Algorithmic bias is a serious challenge that often arises from non-diverse training datasets and reduces accuracy for underrepresented populations. Consequently, proactive subgroup analyses, fairness metrics, and bias mitigation strategies are imperative to ensure equitable AI performance across demographic groups (109).

Model transparency and explainability also pose critical ethical concerns in AI integration into clinical workflows, as many ‘black-box’ algorithms obscure the decision-making process and erode patient-clinician trust. Therefore, adopting explainable AI (XAI) frameworks is essential to foster collaborative clinician-AI decision-making and mitigate overreliance on opaque predictions (110). Importantly, AI can augment clinical judgment; however, oncologists’ decisions must remain central and not be replaced by AI (111).

Future directions

The precise preoperative prediction of ILNM in penile cancer will ultimately rely on integrating advanced multimodal imaging, radiomic features, and molecular/genomic biomarkers. Currently, the progress is hindered by data scarcity, heterogeneity, and a lack of standardized protocols. To overcome these barriers, prospective multicenter collaboration and protocol standardization must be prioritized. Molecular markers of penile cancer are currently in the preliminary stage of exploration. The investigation of non-invasive hematological markers represents a significant avenue for future preoperative research. Additionally, liquid biopsy offers a highly promising non-invasive complement to imaging and pathology. The analysis of circulating tumor DNA (ctDNA), HPV DNA, non-coding RNAs, and extracellular vesicles offers a dynamic method for predicting nodal metastasis and monitoring treatment response in urological cancers (112,113).

The inability of existing predictive approaches to accurately predict ILNM in PSCC underscores the urgent need for advanced AI-driven multimodal predictive techniques. Future research is required to focus on developing and validating AI-driven multimodal models to enhance the accuracy of ILNM prediction in PSCC. Investigating the synergistic potential of combining liquid biopsy biomarkers with AI-enhanced imaging models will be important for advancing early detection, refining surgical decision-making, and longitudinal surveillance in PSCC. Ultimately, prospective, multicenter external validation of these models is a prerequisite for clinical adoption.

Clinical implications

From a clinical perspective, currently available tools for predicting ILNM in PSCC remain limited. Clinicopathological factors and surgical staging techniques such as DSNB represent the most reliable and clinically actionable approaches, albeit with important limitations. Conventional imaging serves a supportive role, while advanced imaging and molecular biomarkers remain adjunctive and investigational.

AI–based and multimodal predictive models show substantial promise for improving individualized risk stratification, but their use should currently be confined to research settings. Until robust prospective validation becomes available, clinical decision-making should continue to rely on established risk factors and guideline-endorsed staging strategies.

Conclusions

Accurate preoperative prediction of ILNM remains a significant challenge in the management of penile SCC. Current predictive models mainly rely on clinicopathological factors and often omit advanced imaging (ADC from DWI) and molecular (HPV/p16, systemic inflammatory) biomarkers. This omission contributes to staging inaccuracies and restricts their clinical utility in ILNM prediction. Owing to data scarcity and a lack of prospective clinical research, current evidence on AI applications for ILNM prediction in penile cancer remains limited. However, this review comprehensively evaluates a wide range of significant predictors of ILNM and provides sufficient literature support. Emerging AI-based techniques can integrate multimodal data into a unified analytical workflow to enhance preoperative ILNM prediction and inform decisions on selective ILND. Clinical translation of these multimodal frameworks is currently limited by scarce and heterogeneous datasets. Therefore, prospective, multicenter studies with external validation are crucial for clinical implementation of AI-based multimodal frameworks in penile cancer.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.