This study included two independent patient cohorts. The Fudan cohort contained 182 patients who underwent surgery for histologically confirmed G1/G2 NF-PanNETs at Fudan University Shanghai Cancer Center (FUSCC), between 2006 and 2018. The external validation set comprised 125 patients with G1/G2 NF-PanNETs who underwent a pancreatectomy at an external hospital with pathology consultations performed at our institute. To study TLS in heterogeneous pancreatic neuroendocrine neoplasms (PanNENs), we enrolled 26 patients with G1/G2 functional PanNETs (22 insulinomas and 4 gastrinomas), 22 patients with G3 NF-PanNETs, and 18 patients with poorly differentiated pancreatic neuroendocrine carcinomas (PanNECs). All 66 patients underwent surgery and their diagnosis was histologically confirmed at FUSCC. Surgical history, clinicopathological parameters, and follow-up information were collected retrospectively from our medical record database. Patients were followed up with every 3 months for the first 2 years after surgery. Then, patients were followed up every 6 months until recurrence or death. At each follow-up, patients underwent a physical examination, laboratory investigation and contrast-enhanced CT or MRI of the chest/abdomen/pelvis. Recurrence-free survival (RFS) was defined as the time from the date of surgery to the date of recurrence. OS was defined as the time from the date of surgery to the date of either death or the last follow-up. Tumors were restaged and regraded based on the 8th edition American Joint Committee on Cancer (AJCC^8th) tumor-node-metastasis (TNM) staging system and 2017 WHO classification.

Formalin-fixed, paraffin-embedded (FFPE) specimens from each patient in the 182-patient Fudan cohort and 66 patients with PanNENs were prepared in 5 µm thick sections for H&E staining.24 H&E-stained sections of the 125-patient external validation set were borrowed from the pathology department at our institute. To further assess lymphocytic organization in G1/G2 NF-PanNETs, immunohistochemistry (IHC) staining was carried out as previously described,25–28 using the 182 serial sections from the Fudan cohort. IHC staining was performed to identify helper T cells (CD4, ab133616, 1:200, Abcam), cytotoxic T cells (CD8, ab172729, 1:300, Abcam), regulatory T cells (Tregs, FOXP3, ab20034, 1:800, Abcam), memory T cells (CD45RO, ab23, 1:1,000, Abcam), B cells (CD20, 24828-1-AP, 1:5,000, Proteintech), dendritic cells (DCs, CD11c, ab52632, 1:500, Abcam), natural killer cells (NK cells, anti-NCR1, ab224703, 1:1,000, Abcam), and tumor-associated macrophages (TAMs, CD68, ab213363, 1:4,000, Abcam).

TLS were identified on full-face H&E slides of 307 tumors from the Fudan cohort, the external validation set, and the 26 G1/G2 functional PanNETs, 22 G3 NF-PanNETs, and 18 PanNECs tissues. Slides were reviewed by two independent observers (Dr Dan Huang and Dr Hai-Ying Zeng) for the presence and the location of TLS. When TLS were identified, the numbers of tumor-infiltrating immune cells outside the TLS were counted in G1/G2 NF-PanNET tissue by evaluating five random high-power fields (×200). Data are presented as the mean±SD.

Five FFPE slides from specimens, with observed TLS in the H&E-stained sections, were subjected to multispectral immunohistochemical (mIHC) staining29 30 using the Opal color kit (PerkinElmer, Hopkinton, Massachusetts, USA), according to the manufacturer’s instructions, to provide simultaneous detection of adaptive immune system cell types, such as CD4⁺ T cells, CD8⁺ T cells, CD20⁺ B cells, and CD45RO⁺ memory T cells. Cell nuclei were stained with DAPI. Multiplexed color slides were imaged with a PerkinElmer Vectra automated multispectral microscope at 100× magnification and analyzed using PerkinElmer inForm Analysis Software.

Nomogram construction used a multivariate Cox model for RFS, based on TLS, WHO classification, and AJCC^8th TNM stage, and was fit to the Fudan cohort. Validation was performed using bootstrap resampling; external validation was performed using the 125-patient external validation set.6 Model performance was evaluated using the concordance index (c-index) and calibration curve.31 The c-index was calculated for the WHO classification and AJCC^8th TNM stage, and the two indices were compared with the c-index from the nomogram.

Statistical analyses were performed using SPSS (V.25.0, IBM, Armonk, New York, USA) and Stata (V.15.0, Stata Corp., Houston, Texas, USA). Nomograms were constructed using R V.3.6.3 (R Foundation for Statistical Computing, Vienna, Austria). Continuous variables were compared by non-parametric tests or Student’s t-test. Categorical variables were compared by Pearson’s Χ² test or Fisher’s test. The Kaplan-Meier method and log-rank test were used to estimate statistical differences in survival. Cox-regression analyses were performed to identify independent prognostic factors of RFS and OS. Variables significantly associated with RFS or OS in the univariate analysis were included in the multivariate analysis. Two-sided p values of <0.05 were considered statistically significant.

The clinicopathological characteristics of two patient cohorts are summarized in table 1. In the Fudan cohort, the 1-year, 3-year, and 5-year recurrence rates were 12.6%, 25.3%, and 28.0%, respectively; and the 1-year, 3-year, and 5-year survival rates were 100%, 96.7%, and 90.7%, respectively. More than half of the patients were women (52.2%). The median age was 53.0 years (range, 25–82 years). The median tumor size was 29.0 mm (range, 6.0–95.0 mm). Positive lymph nodes were present in 36 patients (19.8%). Vascular invasion and perineural invasion were observed in 45 (24.7%) and 39 cases (21.4%), respectively. The percentages of patients with AJCC^8th TNM stages I, II, and III were 39.6%, 40.7%, and 19.8%, respectively, and WHO grades 1 and 2 were 65.9% and 34.1%, respectively.

We defined the presence, location and cellular composition of TLS and related these to the clinicopathological characteristics of G1/G2 NF-PanNETs. TLS were located peritumorally, either just outside the tumor tissue or in the stroma. No TLS were observed in peritumoral normal pancreatic tissues. TLS were variable in size and shape, with most being oval or irregular (online supplemental figure S1), and were observed in 35.7% and 36.8% of G1/G2 NF-PanNETs in the Fudan cohort and external validation set, respectively.

To characterize the cellular composition of TLS, we used IHC to identify CD4⁺ T cells, CD8⁺ T cells, CD20⁺ B cells, CD11c⁺ DCs, CD45RO⁺ memory T cells, anti-NCR1⁺ NK cells, FOXP3⁺ Tregs, and CD68⁺ TAMs in serial sections (figure 1A–H). CD4⁺ and CD8⁺ T cells were primarily in the parafollicular cortex, with CD4⁺ T cells showing more infiltration than CD8⁺ T cells. Most CD20⁺ B cells were in the center of the follicle, whereas CD11c⁺ DCs and CD45RO⁺ memory T cells were located primarily within the T-cell zone, with some dispersion into the center of the follicle. Only sporadic anti-NCR1⁺ NK cells were identified, and CD68⁺ TAMs were rarely observed. FOXP3⁺ Tregs were rare in both TLS and tumor tissues.

Next, we used mIHC to simultaneously detect CD45RO⁺, CD20⁺, CD4⁺, and CD8⁺ cells in G1/G2 NF-PanNET tissues and define their relationship to TLS and the tumor microenvironment (figure 2B–F). We focused on these adaptive immune cells because they were the major components of TLS based on IHC. The mIHC data reflected the IHC results, with CD45RO⁺ memory T cells showing more infiltration and greater dispersion into the center of the follicle in mIHC. Quantification of the cell types in the mIHC images showed that CD45RO⁺ memory T cells were the most common immune cells in TLS, followed by CD20⁺ B cells, CD4⁺ T cells, and CD8⁺ T cells (39.35%, 36.61%, 20.05%, and 3.99%, respectively, figure 2G–H). CD20⁺ B cells tended to be located in the center of the follicle whereas CD4⁺ T cells, CD8⁺ T cells, and CD45RO⁺ memory T cells were primarily located in the parafollicular zone.

Patients with G1/G2 NF-PanNETs containing TLS associated with better clinical outcomes based on clinicopathological characteristics, such as negative lymph node, absence of vascular and perineural invasion, lower TNM stage (I, II), and lower WHO grade (G1), although statistical significance was observed only for vascular invasion. Patients without vascular invasion showed a higher percentage of TLS versus patients with vascular invasion (40.88% vs 20%, p=0.011). There was no significant difference in the percentage of TLS between patients with negative or positive lymph nodes (36.30% vs 33.33%, p=0.739). Likewise, there was no significant difference in the percentage of TLS between patients with or without perineural invasion (30.77% vs 37.06%, p=0.467). Additionally, statistical significance was not reached when comparing the percentage of TLS between patients with AJCC^8th TNM stage I/II or III tumors (36.99% vs 30.56%, p=0.471), and patients with WHO G1 or G2 tumors (40% vs 27.42%, p=0.093, online supplemental figure S2A).

To better characterize TLS in heterogeneous PanNENs, we examined 26 G1/G2 functional PanNETs, 22 G3 NF-PanNETs, and 18 PanNECs. TLS were located peritumorally, varied in shape and size and had similar cellular composition (figure 1I–K). TLS were observed in 38.46% of G1/G2 functional PanNETs, 27.28% of G3 NF-PanNETs, and 11.11% of PanNECs. Although the percentage of TLS in G1/G2 NF-PanNETs was lower than in G1/G2 functional PanNETs and higher than in G3 NF-PanNETs, these differences did not reach statistical significance. However, the percentage of TLS in G1/G2 NF-PanNETs was significantly greater than in PanNECs (p=0.035, online supplemental figure S2B).

Tumor-infiltrating CD4⁺ and CD8⁺ T cells, B cells, and CD45RO⁺ memory T cells are indispensable for forming immune-responsive microenvironment,32–34 whereas TAMs are associated with immune-suppressive microenvironment.12 We identified significantly higher infiltration of CD4⁺ T cells in G1/G2 NF-PanNETs containing more than five TLS (TLS_≥5) compared with those having one to four TLS (TLS_1–4), or no TLS (TLS₀, p=0.031 and p=0.011, respectively). The numbers of tumor-infiltrating CD8⁺ T cells (p=0.032 and p=0.014, respectively) and CD20⁺ B cells (p=0.031 and p<0.001, respectively) were also significantly greater in tumor tissues with TLS_1–4 and TLS_≥5 compared with TLS₀. Additionally, the number of CD45RO⁺ memory T cells was significantly higher in TLS_≥5 G1/G2 NF-PanNETs than TLS₀ G1/G2 NF-PanNETs (p=0.048). In contrast, the number of CD68⁺ TAMs in TLS_1–4 and TLS_≥5 tumors was significantly lower than in TLS₀ tumors (p=0.021 and p=0.007, respectively). Moreover, no significant differences in the numbers of tumor-infiltrating CD11c⁺ DCs and NK cells were uncovered based on TLS number (online supplemental figure S3). Together, these data suggested that TLS might promote immune-responsive microenvironment.

We explored the association of the presence of TLS with RFS and OS in both the Fudan cohort and external validation set to determine whether TLS might be useful in determining patient prognosis. In the Fudan cohort, median RFS was 39 months (range, 1.5–95.0 months) and median OS was 58 months (range, 10.0–96.0 months). The presence of TLS was associated with longer RFS and OS (p<0.001 and p=0.001, respectively, figure 3A,B). In the external cohort, median RFS was 37 months (range, 2.3–128.7 months) and median OS was 56 months (range, 7.6–128.7 months). The presence of TLS was associated with longer RFS and OS (p<0.001 and p<0.001, respectively, figure 3C,D).

Next, we explored the correlation between the number of TLS and patient survival, and found that TLS number (TLS_1–4 vs TLS_≥5) did not significantly impact either RFS or OS; however, patients with any number of TLS exhibited longer RFS and OS than those without TLS (online supplemental table S1).

Multivariate survival analyses indicated that AJCC^8th TNM stage, WHO classification and the presence of TLS were independent prognostic factors of RFS (HR=2.8, 95% CI: 1.79–4.27, p<0.001; HR=2.9, 95% CI: 1.57–5.22, p=0.001 and HR=0.3, 95% CI: 0.14–0.69, p=0.004, respectively) and OS (HR=2.2, 95% CI: 1.14–4.20, p=0.019; HR=3.0, 95% CI: 1.33–6.75, p=0.008 and HR=0.3, 95% CI: 0.12–0.74, p=0.009, respectively) in the Fudan cohort. Likewise, these results were confirmed in the external validation set (table 2).

Median follow-up time was 60 months, and there were 51 recurrences in the 182-patient Fudan cohort and 32 recurrences in the 125-patient external validation set. We aimed to construct a nomogram, incorporating the presence of TLS, to predict the probability of 5-year RFS, based on the multivariate Cox model (figure 4A). The nomogram c-index on the Fudan cohort was 0.772, superior to c-indices based on the WHO classification (0.665) and AJCC^8th TNM stage (0.720). The nomogram c-index on the external validation set was 0.879, which was also superior to that based on the WHO classification (0.777) and AJCC^8th TNM staging (0. 807). The calibration plots for the Fudan cohort and external validation set showed good agreement between the predicted and observed 5-year recurrence-free probabilities (figure 4B,C).

Memorial Sloan-Kettering Cancer Center (MSKCC) constructed a nomogram using four variables: number of positive lymph nodes, Ki67 index, tumor diameter, and presence of vascular or perineural invasion.6 For comparison, we constructed a MSKCC-based nomogram using these same variables and applied the nomogram to the Fudan cohort and external validation set (figure 4D).6 The c-indices of the MSKCC-based nomogram on the Fudan cohort and external validation set were 0.692 and 0.719, respectively, indicating less discrimination than the new nomogram based on TLS, WHO classification and AJCC^8th TNM stage. Calibration plots for the two cohorts were shown in figure 4E,F.