Deep learning for predicting pituitary neuroendocrine tumour lineage and high-risk subtypes from histology

Abstract

Pituitary neuroendocrine tumours (PitNETs) exhibit significant heterogeneity, posing challenges for clinical management. We developed a deep learning model to predict PitNET lineage, high-risk subtypes, and recurrence directly from routine H&E-stained whole-slide images. Trained on 925 patients from USTC and externally validated on cohorts from Taihe Hospital (n = 226) and Huashan Hospital (n = 193), the model achieved a micro-average AUC of 0.912 for lineage classification (SF1: 0.926, PIT1: 0.932, TPIT: 0.904; Without distinct lineage: 0.706). High-risk subtype prediction yielded AUCs of 0.805 (PIT1), 0.753 (TPIT), and 0.733 (null cell). Recurrence prediction reached an AUC of 0.641. Analysis of the tumour microenvironment revealed that compared with primary tumours, recurrence tumours were characterized by an increased density of M2 macrophages and decreased infiltration of CD8 + T cells. Spatial transcriptomics further elucidated distinct molecular pathways associated with recurrence, providing mechanistic insights into prognostic predictions. Our deep learning model accurately predicts PitNET characteristics from routine H&E slides, and spatial biology validation identified distinct immune and molecular features associated with recurrence.

Introduction

PitNETs are neuroendocrine tumours that arise from the adenohypophyseal cells of the anterior pituitary and are characterized by neuroendocrine granules in their ultrastructure^1,2. They represent approximately 10–20% of all primary intracranial tumours, making them the third most common type after gliomas and meningiomas³. In line with the revised 5th WHO guidelines promoting a standardized classification system for neuroendocrine neoplasms, the dual designation “PitNET/pituitary adenoma” has been introduced to describe tumours originating from adenohypophyseal cells, with “PitNETs” now established as the preferred diagnostic nomenclature⁴. Beyond their mass effect, PitNETs significantly impact pituitary gland function, resulting in either hormonal hypersecretion or hyposecretion, which subsequently triggers a cascade of systemic endocrine and metabolic dysfunctions⁵.

The 5th WHO classification system not only underscores the nomenclature transition to PitNETs, but also advocates for an enhanced classification framework that incorporates transcription factors as pivotal markers for identifying cell differentiation lineages on the basis of adenohypophyseal hormone secretion patterns demonstrated through immunohistochemical analysis⁶. On the basis of their cellular differentiation lineages, PitNETs are systematically classified into four principal lineages encompassing sixteen distinct subtypes: the SF1 lineage, the PIT1 lineage, the TPIT lineage, and PitNETs without distinct lineages⁷. Moreover, in light of the varying clinical prognoses and distinct invasive biological behaviours exhibited by different PitNET subtypes, the diagnostic category of “high-risk PitNETs” has been established to identify tumours with aggressive clinical features⁸. This category encompasses several clinically aggressive variants, notably sparsely granulated somatotroph tumours, immature plurihormonal PIT1 lineage tumours, and TPIT lineage-derived Crooke cell tumours. These high-risk subtypes typically present as macroadenomas exceeding 10 mm in diameter, demonstrating invasive growth patterns and showing limited therapeutic response to pharmacological interventions, ultimately resulting in a poor clinical prognosis^9–11. Consequently, these high-risk subtypes necessitate rigorous clinical surveillance and prompt, aggressive therapeutic management to optimize patient outcomes. Currently, the pathological classification of PitNETs predominantly depends on comprehensive immunohistochemical (IHC) antibody panels; however, this approach is significantly limited by multiple inherent drawbacks, including procedural complexity, time-intensive protocols, challenges in quality control standardization, and substantial financial burden. Consequently, there is an urgent need to develop and implement novel, objective diagnostic methods that can accurately identify differentiation lineages and reliably predict high-risk subtypes of PitNETs, thereby enhancing both diagnostic precision and clinical efficiency.

Haematoxylin and eosin (H&E) staining remains the cornerstone of routine histopathological diagnosis, providing comprehensive morphological information about tissue architecture and cellular characteristics that are fundamental for accurate diagnostic interpretation. The integration of digital pathology workflows with advanced data-driven artificial intelligence (AI), particularly deep learning algorithms, has enabled the development of numerous computational models capable of predicting molecular-level alterations in other types of neoplasms through in-depth analysis of H&E-stained histopathological images^12–14. However, current research on predicting PitNETs from H&E slides remains limited.

In this study, we developed and rigorously validated a novel computational framework that enables rapid, accurate, and objective identification of differentiation lineages and high-risk subtypes of PitNETs, thereby establishing a robust diagnostic foundation for optimizing therapeutic decision-making in PitNET management. Moreover, we implemented a comprehensive postoperative surveillance protocol for PitNET patients, employing H&E-stained whole slide images (WSIs) for recurrence prediction. To further enhance the interpretability of WSI-based recurrence prediction, we conducted advanced molecular profiling through multiplex immunofluorescence staining and spatial transcriptomic analysis of matched primary and recurrence tumour specimens.

Results

Clinicopathological features of the study cohorts

The USTC cohort included 925 PitNET patients (median age 46; 471 males, 454 females), and the clinicopathologic details are shown in Table 1. Patients with functional tumours (56.5%, 523 patients) presented with hormonal symptoms (e.g., limb hypertrophy and lactation/menstrual changes), whereas those with nonfunctioning tumours (43.5%, 402 patients) presented with mass effect symptoms (e.g., visual impairment and headache). The distribution of tumour sizes was as follows: microadenomas (<10 mm), 13.2%; macroadenomas (10–39 mm), 83.1%; and giant adenomas (≥40 mm), 3.7%. According to the 5th WHO classification diagnostic criteria, there were 449 cases of the SF1 lineage (48.54%), 264 cases of the PIT1 lineage (28.54%), 140 cases of the TPIT lineage (15.14%), and 72 cases of PitNETs without distinct lineages (7.78%). Among the 925 cases analysed, the distribution of seven identified high-risk subtypes was as follows: male lactotroph tumours (MLT, n = 40, 4.32%), sparsely granulated somatotroph tumours (SGST, n = 31, 3.35%), acidophil stem cell tumours (ASCT, n = 11, 1.19%), immature plurihormonal PIT1-lineage tumours (IPPT, n = 24, 2.59%), silent corticotroph tumours (SCT, n = 58, 6.27%), Crooke cell tumours (CCT, n = 6, 0.65%), and null cell tumours (NCT, n = 27, 2.92%). During follow-up surveillance, tumour recurrence occurred in 97 patients (10.48%), which was further stratified as follows: 76 patients (78.4%) experienced histologically confirmed recurrence requiring secondary or tertiary surgical intervention, while 21 patients (21.6%) exhibited radiological evidence of recurrence. Importantly, complete surgical documentation for both primary and recurrence procedures was available for all 32 recurrence cases at our institution, enabling paired histopathological analysis.

Table 1.

Clinicopathologic features in the USTC cohort

Lineage	N	Gender		Tumour size
		male	female	microadenomas	macroadenomas	giant adenomas
SF1 Lineage
Gonadotroph tumour	449 (48.54%)	306 (68.15%)	143 (31.85%)	30 (6.68%)	406 (90.42%)	13 (2.90%)
PIT1 Lineage	264 (28.54%)	123 (46.59%)	141 (53.41%)	51 (19.32%)	199 (75.38%)	14 (5.30%)
Densely granulated somatotroph tumour	29 (3.14%)	16 (55.17%)	13 (44.83%)	5 (17.24%)	23 (79.31%)	1 (3.45%)
Sparsely granulated somatotroph tumour	31 (3.35%)	8 (25.81%)	23 (74.19%)	2 (6.45%)	27 (87.10%)	2 (6.45%)
Mammosomatotroph tumour	73 (7.89%)	37 (50.68%)	36 (49.32%)	13 (17.81%)	58 (79.45%)	2 (2.74%)
Mixed somatotroph-lactotroph tumour	15 (1.62%)	8 (53.33%)	7 (46.67%)	4 (26.67%)	9 (60.0%)	2 (13.33%)
Densely granulated lactotroph tumour	49 (5.30%)	30 (61.22%)	19 (38.78%)	15 (30.61%)	29 (59.18%)	5 (10.21%)
Sparsely granulated lactotroph tumour	9 (0.97%)	4 (44.44%)	5 (55.56%)	1 (11.11%)	6 (66.67%)	2 (22.22%)
Acidophil stem cell tumour	11 (1.19%)	4 (36.36%)	7 (63.64%)	6 (54.55%)	5 (45.45%)	0 (0%)
Thyrotroph tumour	3 (0.32%)	1 (33.33%)	2 (66.67%)	0 (0%)	3 (100%)	0 (0%)
Mature plurihormonal PIT1 lineage tumour	20 (2.16%)	6 (30.0%)	14 (70.0%)	2 (10.0%)	18 (90.0%)	0 (0%)
Immature plurihormonal PIT1 lineage tumour	24 (2.59%)	9 (37.5%)	15 (62.5%)	3 (12.5%)	21 (87.5%)	0 (0%)
TPIT Lineage	140 (15.14%)	19 (13.57%)	121 (86.43%)	25 (17.86%)	109 (77.86%)	6 (4.28%)
Densely granulated corticotroph tumour	31 (3.35%)	10 (32.26%)	21 (67.74%)	11 (35.48%)	20 (64.52%)	0 (0%)
Sparsely granulated corticotroph tumour	103 (11.14%)	7 (67.96%)	96 (93.20%)	13 (12.62%)	84 (81.55%)	6 (58.25%)
Crooke cell tumour	6 (0.65%)	2 (33.33%)	4 (66.67%)	1 (16.67%)	5 (83.33%)	0 (0%)
PitNETs without distinct lineage	72 (7.78%)	23 (31.94%)	49 (68.06%)	16 (22.22%)	55 (76.39%)	1 (1.39%)
Plurihormonal tumour	45 (4.86%)	13 (28.89%)	32 (71.11%)	12 (26.67%)	32 (71.11%)	1 (2.22%)
Null cell tumour	27 (2.92%)	10 (37.04%)	17 (62.96%)	4 (14.81%)	23 (85.19%)	0 (0%)

For the external validation, the Taihe Hospital cohort comprised 226 patients (119 males, 107 females) with the following lineage distributions: SF1 lineage (n = 90, 39.82%), PIT1 lineage (n = 85, 37.61%), TPIT lineage (n = 36, 15.93%), and PitNETs without distinct lineages (n = 15, 6.64%). The Huashan Hospital cohort included 193 patients (122 males, 71 females) with comparable lineage distribution patterns: SF1 lineage (n = 76, 39.38%), PIT1 lineage (n = 58, 30.05%), TPIT lineage (n = 49, 25.39%), and PitNETs without distinct lineages (n = 10, 5.18%). The detailed clinicopathological characteristics of both validation cohorts are presented in Supplementary Tables 1, 2.

Deep learning-based prediction of PitNETs

For internal validation, the performance of the established model was evaluated on 925 cases with 989 WSIs comprising four lineages: the SF1 lineage (n = 449), PIT1 lineage (n = 264), TPIT lineage (n = 140), and WDL (n = 72). The model achieved patient-level areas under the receiver operating characteristic (ROC) curve (AUCs) of 0.926 (95% CI: 0.885–0.959) for the SF1 lineage, 0.932 (95% CI: 0.892–0.966) for the PIT1 lineage, 0.904 (95% CI: 0.851–0.952) for the TPIT lineage and 0.706 (95% CI: 0.554–0.859) for the WDL group (Fig. 1A–B and Table 2). The model demonstrated the following performance metrics for SF1 lineage prediction: accuracy (ACC) 84.6%, sensitivity 83.9%, specificity 85.3%, positive predictive value (PPV) 84.8%, and F1 score 0.843. For PIT1 lineage prediction, the corresponding metrics were an ACC of 87.8%, a sensitivity of 66.0%, a specificity of 96.3%, a PPV of 87.5%, and an F1 score of 0.752. The complete performance metrics, including the ACC, sensitivity, specificity, PPV, and F1 score, for the other lineages are detailed in Table 2. For external validation, 226 patients were included in the Taihe Hospital cohort, and 193 patients were included in the Huashan Hospital cohort. As shown in Table 2, the model achieved prediction AUC values of 0.830 and 0.860 across the four lineages (Fig. 1C–F and Table 2). External validation of independent cohorts from the Taihe Hospital cohort and Huashan Hospital cohort predominantly demonstrated the generalizability of this model. To demonstrate the superiority of our DTransMIL model, we compared it against the TransMIL and ACMIL models. In the cohort of 925 cases from USTC, compared with the TransMIL (Fig. 1G) and ACMIL models (Fig. 1H), the DTransMIL model achieved significantly higher AUC values. The specific AUC values for all three models are presented as a bar chart in Fig. 1I, which confirms that the observed differences are statistically significant for the SF1, PIT1, and TPIT lineages.

Fig. 1. ROC curve and classification confusion matrix of PitNETs in each cohort.

A, B ROC curve and classification confusion matrix of predicting lineages classification in USTC cohort. C–F ROC curve and classification confusion matrix of predicting lineage classification in Taihe cohort and Huashan cohort. G, H ROC curve of predicting lineages classification in USTC cohort using TransMIL and ACMIL. I Bar chart of the AUC values for the three models. J–O ROC curve and classification confusion matrix of predicting high-risk types in PIT1 lineage, TPIT lineage and WDL. P–Q ROC curve and classification confusion matrix of predicting recurrence. MLT male lactotroph tumours, SGST sparsely granulated somatotroph tumours, ASCT acidophil stem cell tumours, IPPT immature plurihormonal PIT1-lineage tumours, SCT silent corticotroph tumours, CCT Crooke cell tumours, NCT null cell tumours.

Table 2.

Classification results on internal and external test sets

	Lineage	AUC (95%CI)	ACC	Sensitivity	Specificity	PPV	F1 score
USTC cohort	SF1 lineage	0.926 (0.885–0.959)	84.6% (159/188)	83.9% (78/93)	85.3% (81/95)	84.8% (78/92)	0.843
	PIT1 lineage	0.932 (0.892–0.966)	87.8% (165/188)	66.0% (35/53)	96.3% (130/135)	87.5% (35/40)	0.752
	TPIT lineage	0.904 (0.851–0.952)	89.4% (168/188)	67.9% (19/28)	93.1% (149/160)	63.3% (19/30)	0.655
	Without distinct lineage	0.706 (0.554–0.859)	86.2% (162/188)	50.0% (7/14)	89.1% (155/174)	27.0% (7/26)	0.351
Taihe cohort	SF1 lineage	0.822 (0.738–0.903)	75.5% (77/102)	76.9% (30/39)	74.6% (47/63)	65.2% (30/46)	0.706
	PIT1 lineage	0.766 (0.664–0.861)	73.5% (75/102)	61.5% (24/39)	81.0% (51/63)	66.7% (24/36)	0.640
	TPIT lineage	0.801 (0.688–0.896)	82.4% (84/102)	44.4% (8/18)	90.5% (76/84)	50.0% (8/16)	0.470
	Without distinct lineage	0.728 (0.493–0.942)	94.1% (96/102)	42.9% (3/7)	97.9% (93/95)	60.0% (3/5)	0.500
Huashan cohort	SF1 lineage	0.867 (0.785–0.932)	78.7% (74/94)	84.2% (32/38)	75.0% (42/56)	70.0% (32/46)	0.764
	PIT lineage	0.748 (0.645–0.839)	73.4% (69/94)	56.7% (17/30)	81.3% (52/64)	58.6% (17/29)	0.576
	TPIT lineage	0.863 (0.768–0.940)	85.1% (80/94)	57.1% (12/21)	93.2% (68/73)	70.1% (12/17)	0.630
	Without distinct lineage	0.681 (0.254–1.000)	96.8% (91/94)	40.0% (2/5)	100.0% (89/89)	100.0% (2/2)	0.571

In the prediction of the five high-risk subtypes within the PIT1 lineage, the model achieved an AUC of 0.805, with the highest AUC (0.853) observed for male lactotroph tumours. For the two high-risk subtypes in the TPIT lineage, the overall AUC was 0.753; however, the Crooke cell subtype had a markedly lower AUC (0.313), likely attributable to its limited sample size (only 6 cases among the 140 cases of the TPIT lineage). In the prediction of high-risk subtypes without distinct lineages, the model yielded an AUC of 0.733. Additionally, for predicting PitNET recurrence, the model achieved an AUC of 0.641 (Fig. 1J–Q and Supplementary Table 3).

Model interpretability

To validate model interpretability, we generated attention heatmaps quantifying the region-specific focus. These maps effectively distinguished cellular morphologies across four lineages and seven high-risk subtypes by identifying chromophobic/oncocytic, eosinophilic, and basophilic cytoplasmic features in H&E-stained WSIs. As shown in Fig. 2, tumour cells of the SF1 lineage typically exhibited a chromophobic or oncocytic morphology and display polygonal to elongated shapes with distinct cellular polarity. While diffuse and sinusoidal growth patterns may occur, the most consistent and diagnostically significant histological features included a characteristic trabecular architecture, perivascular pseudorosettes, and, occasionally, papillary structures. In the PIT1 lineage, tumours consisted of cells with abundant eosinophilic (acidophilic) cytoplasm. TPIT-lineage corticotroph tumours, by contrast, contained variable proportions of basophilic cells. Among PitNETs without distinct lineages, null cell tumours were predominantly composed of chromophobic or basophilic cells, while plurihormonal tumours demonstrated heterogeneous cellular morphology reflecting mixed lineage characteristics. The seven high-risk PitNET subtypes also exhibited characteristic histological patterns (Fig. 3), such as (1) acidophil stem cell tumours characterized by cytologically immature cells with a chromophobic to mildly acidophilic cytoplasm, containing clear vacuoles resulting from mitochondrial aggregation, and (2) Crooke cell tumours composed of cells with translucent cytoplasm and centrally located nuclei.

Fig. 2. Heatmaps for predicting PitNETs lineages classification interpretation.

A Low-power view of HE-stained sections. B Low-power view of heatmaps. C High-power view of heatmaps. D High-power view of HE-stained sections(×400). E Diagnostic IHC of various lineages. WDL without distinct lineage.

Fig. 3. Heatmaps for predicting PitNETs high-risk types interpretation.

A Low-power view of HE-stained sections. B Low-power view of heatmaps. C High-power view of heatmaps. D High-power view of HE-stained sections (×400). E Diagnostic IHC of various high-risk subtypes. MLT male lactotroph tumours, SGST sparsely granulated somatotroph tumours, ASCT acidophil stem cell tumours, IPPT immature plurihormonal PIT1-lineage tumours, SCT silent corticotroph tumours, CCT Crooke cell tumours, NCT null cell tumours.

Tumour microenvironment analysis

PitNETs show minimal morphological variation across lineages and subtypes, yet primary and recurrence tumours within the same category appear highly similar. To predict AI-based recurrence, we investigated the tumour immune microenvironment (TIME), a known driver of PitNET progression. Previous studies report significant differences in the TIME between primary and recurrence PitNETs, suggesting its potential role in recurrence^15,16. To investigate whether artificial intelligence can identify these distinct TIME patterns in both primary and recurrence tumours, we employed two multiplex immunofluorescence (mIF) panels to characterize the differences in the immune microenvironment between primary and recurrence tumours. As depicted in Fig. 4A, in both primary and recurrence cases of the SF1 lineage, PIT1 lineage, and TPIT lineage, we found that the immune microenvironment cells of the recurrence cases were significantly different from those of the primary cases. In recurrence cases, the number of CD3 + CD4 + T lymphocytes was significantly greater than that in primary cases, the number of CD3 + CD8 + T lymphocytes was significantly lower than that in primary cases, and the CD3 + CD8 + /CD3 + CD4+ ratio was even lower in recurrence cases. In the recurrence cases, the number of M2 macrophages positive for CD68 + CD163 was greater than that in the primary cases, the CD68 + CD163 + /CD68+ ratio was also greater than that in the primary cases, the number of neutrophils positive for CD11b was lower than that in the primary cases, and there was no significant difference in B lymphocytes. Spatial analysis revealed that CD3 + CD8 + T lymphocytes were farther from tumour cells, whereas CD4 + T lymphocytes were closer to tumour cells in the recurrence cases. The results of the statistical analyses demonstrating differences in the quantitative and spatial distributions of these immune cells are presented in Fig. 4B. In the PitNETs without distinct lineages, no significantly different immune microenvironmental phenotypes were found between primary and recurrence cases, probably because the number of matchable cases was too low. The quantitative immune markers of paired primary cases and recurrence cases are shown in Table 3.

Fig. 4. Tumour microenvironment analysis in primary and recurrence cases.

A Panel 1 included CD3, CD4, CD8, PD-1, FOXP3 and Syn, and Panel 2 included CD68, HLA-DR, CD163, CD11b, CD20 and Syn. B The statistical analyses demonstrated quantitative and spatial distribution differences of these immune cells in primary and recurrence cases. The y-axis represents immune cell density (average cells/mm²). RT recurrence tumour, PT primary tumour, TCs tumour cells.

Table 3.

Lymphocytes infiltration and immune-to-tumour cell distances between primary and recurrence tumour

Immune markers	SF1 lineage		PIT1 lineage		TPIT lineage		PitNETs without distinct lineage
	PT	RT	PT	RT	PT	RT	PT	RT
CD3 + (average,cells/mm2)	18.66	122.01	76.69	113.48	95.39	206.64	41.65	80.89
CD3 + CD4 + (average,cells/mm2)	9.98	30.00	34.76	58.64	25.91	48.49	7.13	17.44
CD3 + CD8 + (average,cells/mm2)	16.43	0.47	10.84	0.11	7.71	1.69	5.55	2.34
CD3 + CD8/CD3 + CD4 Ratio (%)	0.88	0.02	0.42	0.0038	0.43	0.23	0.62	0.12
CD3+FoxP3 + (average,cells/mm2)	2.19	0.39	3.53	1.30	9.13	3.08	5.88	4.34
CD3 + PD-1 + (average,cells/mm2)	1.29	4.89	42.05	14.61	4.47	43.84	20.20	27.80
CD11b + (average,cells/mm2)	343.13	57.95	314.52	46.59	382.05	85.31	114.46	82.42
CD20 + (average,cells/mm2)	13.17	192.91	107.56	133.91	350.64	83.50	154.44	199.39
CD68 + (average,cells/mm2)	376.17	415.64	36.95	90.05	199.38	225.36	71.93	251.92
CD68 + CD163 + (average,cells/mm2)	23.56	168.96	22.64	50.66	62.67	145.40	8.38	15.68
CD68 + HLA-DR + (average,cells/mm2)	56.67	181.77	13.84	60.95	57.36	67.56	37.65	66.21
CD68 + CD163 + /CD68+ Ratio (%)	0.055	0.32	0.45	0.59	0.16	0.61	0.19	0.22
CD3 + CD4+ to TCs (average,μm)	15.99	10.27	65.27	21.67	35.98	11.84	33.79	17.61
CD3 + CD8+ to TCs (average,μm)	9.309	35.98	22.05	64.16	27.06	48.99	48.34	69.53
CD3+FoxP3+ to TCs (average,μm)	16.979	11.77	27.41	79.37	47.17	26.84	62.67	65.72
CD3 + PD-1+ to TCs (average,μm)	9.66	7.32	23.08	121.45	4.18	29.96	50.79	47.13
CD68 + CD163+ to TCs (average,μm)	26.08	17.17	11.23	36.57	11.19	23.69	16.07	41.05
CD68 + HLA-DR+ to TCs (average,μm)	25.55	18.73	10.23	29.93	11.41	23.81	17.30	27.56

RT recurrence tumour, PT primary tumour, TC tumour cell

Spatial transcriptome analysis

To further investigate the differences in gene expression profiles and relevant signalling pathways between primary and recurrence PitNETs, we performed spatial transcriptomic sequencing on six matched primary–recurrence PitNET pairs. The overall histological structure of the tissue microarray (TMA) samples is shown in Fig. 5A, with spatial transcriptomic analysis revealing the distribution and dynamic interactions of 23 identified cell types (e.g., tumour cells, immune cells, and cancer-associated fibroblasts). Comparative analysis of gene expression profiles between recurrence and primary cases revealed significant transcriptomic reprogramming (Fig. 5B), with 1248 upregulated and 227 downregulated genes in recurrence cases. The significantly upregulated genes included DDX3Y, ARHGAP36, SREBF1, FGFR3, PKD1, and SV2C, whereas the downregulated genes included FAM107A, EPAS1 and PDGFD, suggesting metabolic, angiogenic and neuroendocrine alterations in recurrence tumours.

Fig. 5. Spatial transcriptome analysis in primary and recurrence cases.

A Spatial transcriptomic analysis revealing the distribution and dynamic interactions of 23 identified cell types. B Comparative analysis of gene expression profiles between recurrence and primary cases. C RT-qPCR showed that DDX3Y, ARHGAP36, SREBF1, FGFR3, PKD1, and SV2C were significantly upregulated in the recurrence group, whereas FAM107A, EPAS1, and PDGFD were significantly downregulated. D TIME analysis between recurrence and primary cases. E Primary and recurrence cases exhibited distinct tumour-immune cell interaction patterns. F, G Ligand-receptor interaction analysis showed significant differences in receptor expression profiles between groups. H GO enrichment analysis of upregulated genes in recurrence cases.

We also performed orthogonal validation using RT-qPCR on clinical samples (n = 12; 6 primary and paired recurrence samples). The results showed that the expression trends of these nine genes were highly consistent with the spatial transcriptomics data (Fig. 5C): DDX3Y, ARHGAP36, SREBF1, FGFR3, PKD1, and SV2C were significantly upregulated in RT, while FAM107A, EPAS1, and PDGFD were significantly downregulated (p < 0.05).

With respect to the mIF results, we focused our TIME analysis on both primary and recurrence cases. TIME analysis revealed increased infiltration of M2 macrophages (CD163 + ) around tumour cells, along with more CD4 + T cells and fewer CD8 + T cells in recurrence cases (Fig. 5D), which is consistent with previous multiplex immunofluorescence (mIF) results and confirms an immunosuppressive TIME. Primary and recurrence cases also exhibited distinct tumour–immune cell interaction patterns (Fig. 5E), suggesting that immune evasion occurs through altered cell–cell communication. Ligand–receptor interaction analysis (CellPhoneDB) revealed significant differences in receptor expression profiles between groups (Fig. 5F, G). GO enrichment analysis of upregulated genes in recurrence patients (Fig. 5H) revealed immune regulation, metabolic reprogramming, and neuroendocrine functions during recurrence through immunosuppressive cell enrichment and metabolic adaptation, which may promote tumour recurrence and progression.

Discussion

PitNETs exhibit substantial heterogeneity in terms of biological behaviour and clinical outcomes, necessitating precise diagnostic and prognostic tools^17,18. In this study, we developed an artificial intelligence-based multi-instance learning (MIL) model, called DTransMIL, for the prediction of lineages and high-risk subtypes and the recurrence of patients with PitNETs based on WSIs. During the model development process, a scaled-up magnification workflow was employed. This approach mimics how pathologists examine histopathological images and utilizes attention-based techniques to focus on individual cellular regions without requiring manual annotation. Our model demonstrated robust performance, particularly in terms of lineage classification (AUC = 0.912), while also providing clinically meaningful predictions for high-risk subtypes and recurrence. These findings suggest that computational pathology can complement and potentially streamline traditional diagnostic workflows, reducing the reliance on extensive IHC profiling while maintaining accuracy.

Currently, AI has been widely applied in the pathological diagnosis of various malignancies, including breast cancer, prostate cancer, gastric cancer, colorectal cancer, and glioma^19–23. Its roles in assisting pathological diagnosis and treatment primarily include screening and diagnosis, predicting treatment response, and assessing prognosis. AI research on PitNETs focuses mainly on predicting recurrence, drug responsiveness, and risk stratification. One study developed a multivariate model combining seven machine learning algorithms with 17 clinical parameters from 354 Cushing’s disease patients to predict postoperative recurrence following transsphenoidal surgery²⁴. Another investigation employing gradient-boosting decision trees achieved accurate prediction of surgical remission rates in a cohort of 668 acromegaly patients²⁵. For treatment response forecasting, research utilizing AI-based high-dimensional texture analysis reported 85.1% accuracy in predicting somatostatin analogue efficacy among 47 acromegaly cases²⁶. Innovative work applying deep learning to facial image analysis demonstrated strong predictive performance using data from 814 Cushing’s disease patients, 1,131 acromegaly patients, and 12,598 healthy controls²⁷. In digital pathology, a recent study established an objective classification system for silent PIT1 lineage PitNETs through computational image analysis²⁸. Our study developed and trained a deep learning model for predicting the lineage, high-risk subtypes, and recurrence of PitNETs. The trained model reliably predicted tumour lineages in the test cohort. This approach overcomes the limitations of current diagnostic methods for PitNETs, which require extensive and costly immunohistochemical antibody panels, while demonstrating excellent predictive performance. The model achieved an area under the curve (AUC) of 0.912 in the internal validation cohort, with AUCs of 0.830 and 0.860 in the two external validation cohorts.

Our model delivers three key advantages while maintaining interpretability. First, the novel DTransMIL architecture uniquely integrates macro- and microscale feature extraction, mimicking pathological diagnosis by combining tissue-level patterns with cellular details. This approach achieved superior lineage prediction performance, significantly outperforming the TransMIL and ACMIL models on 925 USTC cases. Second, we efficiently repurpose deep features for high-risk subtype and recurrence prediction using SVM classifiers. The robust preprocessing pipeline includes automated tissue segmentation and data augmentation, whereas our PLIP-guided attention automatically identifies diagnostically relevant regions. Third, shared feature extraction across tasks enhances computational efficiency. Combining deep learning with traditional methods, our unified system addresses lineage classification, risk stratification, and prognosis prediction.

The lack of interpretability associated with the “black box” nature of neural networks²⁹ has become a major limitation hindering the clinical application of deep learning models. In this study, we employed visualization heatmaps to demonstrate that the model is able to automatically locate regions crucial for lineage prediction from WSIs and extract informative features from these regions. To enhance model interpretability, we integrated TIME profiling and transcriptomic analysis. The role of the TIME has been widely acknowledged as a significant contributor to the growth and advancement of PitNETs^30–32. Among the immune cells infiltrating PitNETs, macrophages and T lymphocytes are the most commonly observed³³. Additionally, other immune cell types, including B lymphocytes, FOXP3-positive cells, neutrophils, natural killer (NK) cells, and mast cells, may also be found, although they appear less frequently²². The increased infiltration of M2-type macrophages surrounding tumour cells, along with increased CD4 + T lymphocytes and exhaustion of CD8 + T lymphocytes, suggests a poor prognosis^30,34. In our study, recurrence tumours exhibited an immunosuppressive TIME characterized by increased M2 macrophages and decreased CD8 + T cells. Spatial analysis revealed that CD3 + CD8 + T lymphocytes were located farther from tumour cells, whereas CD4 + T lymphocytes were in closer proximity, which is consistent with the established mechanisms of immune evasion in aggressive PitNETs.

Spatial transcriptomics further elucidated distinct molecular pathways associated with recurrence, providing biological validation for AI-derived risk stratification. Through comparative analysis of gene expression profiles between recurrence and primary tumour cases, we identified the following significantly upregulated genes with established oncogenic relevance: DDX3Y, which drives tumour cell proliferation through Y chromosome-related functions³⁵; ARHGAP36, a Rho GTPase signalling regulator that promotes invasive phenotypes³⁶; and SREBF1, a master lipid metabolism regulator implicated in tumour metabolic reprogramming³⁷. The key signalling molecules FGFR3 and PKD1³⁸ were similarly elevated. The expression of the synaptic vesicle protein SV2C, which potentially regulates neuroendocrine secretion, was also markedly upregulated. Conversely, significantly downregulated genes included tumour suppressors such as FAM107A, the hypoxia-inducible factor EPAS1, and the proangiogenic factor PDGFD, suggesting metabolic, angiogenic and neuroendocrine alterations in recurrence tumours^39,40. We also performed orthogonal validation using RT-qPCR on clinical samples, and the results were highly consistent with the spatial transcriptomics analysis. Furthermore, distinct tumour–immune cell interaction patterns and differential receptor expression profiles between the primary and recurrence groups indicated potential immune evasion through altered cell–cell communication and immune cell subset regulation. These findings collectively demonstrate tumour microenvironment remodelling during recurrence, characterized by enrichment of immunosuppressive cells (M2 macrophages and Tregs) and T-cell exhaustion, which may collectively drive tumour recurrence and progression. AI-based approaches may enable the identification of these microenvironmental features, making recurrence prediction feasible.

This study has several limitations. First, although cases from three medical institutions were included, the sample size remains relatively limited. Specifically, the numbers of high-risk subtypes and recurrence cases from Taihe Hospital and Huashan Hospital were insufficient for adequate external validation. Notably, not a single case of Crooke cell tumour was available from these two institutions. Additional multicentre data are needed to increase the robustness of the AI model, particularly for predicting high-risk subtypes and recurrence. Second, regarding model interpretability, the limited availability of paired cases with both multiplex immunofluorescence and spatial transcriptomic sequencing data precluded separate comparative analyses between primary and recurrence cases across different tumour lineages. Third, due to the limited number of cases in PIT1 lineage that received preoperative medical treatment, it is currently not possible to make a statistically reliable assessment of whether drug-induced histological changes systematically affect the model’s classification performance. Finally, the development of this model requires the involvement of pathologists with specialized expertise in this field to ensure accurate interpretation of the data.

In conclusion, our deep learning framework demonstrates that H&E-based computational pathology can effectively predict PitNET lineage, high-risk subtypes, and recurrence risk. By combining AI with spatial multiomics, we provide a biologically interpretable tool that connects histomorphology and molecular pathology, offering a step towards precision medicine in PitNET management.

Methods

Study population

This multicentre study analysed data from three hospitals, USTC (925 cases, divided into 740 training and 185 testing cases), along with 226 cases from Taihe Hospital and 193 cases from Huashan Hospital as external validation sets. The inclusion criteria were as follows: (1) histopathologically confirmed PitNETs according to the 5th WHO classification of endocrine and neuroendocrine tumours; (2) exclusion of cases with indeterminate immunohistochemical classification; and (3) availability of complete clinicopathological data and follow-up records. All H&E slides were digitally scanned (Sqray SQS-600P) in SVS format. Cases with suboptimal quality or inadequate features were excluded. Two senior pathologists independently reviewed all the cases, and only diagnostically concordant samples were retained for analysis. This study was conducted in strict accordance with the Declaration of Helsinki. The study protocol was reviewed and approved by the Ethics Committees of the First Affiliated Hospital of USTC (Approval No. 2024-RE-164), Huashan Hospital (Approval No. KY2025-1172), and Taihe Hospital (Approval No. 2025KS133). Given the retrospective nature of this study and the use of anonymized data, the requirement for informed consent was waived by the approving ethics committees.

Computational framework

The overall framework is illustrated in Fig. 6A. We propose a deep-learning pipeline to predict PitNET lineages, high-risk subtypes, and recurrence using only H&E-stained WSIs. Its core, detailed in Fig. 6B, is a dual-branch transformer-based MIL (DTransMIL) model. A macrobranch locates critical tissue patches, guiding a microbranch for detailed feature extraction. The combination of features enables accurate lineage prediction. For high-risk subtype classification and recurrence prediction, we introduce dedicated modules (HRSC and RP). Both leverage the deep features from the DTransMIL model for risk assessment. This integrated pipeline shows great potential for enhancing PitNET diagnosis and prognosis, aiding clinical decision-making.

Fig. 6. Computational framework for predicting PitNETs analysis using WSIs.

A Lineage prediction module: Our DTransMIL framework employs a dual-branch Transformer architecture for accurate lineage determination. The macro-branch first identifies diagnostically relevant regions through attention mechanisms, which then guide the micro-branch’s fine-grained feature extraction. The integrated multi-scale features enable precise classification of PitNET lineages. B Prognostic prediction module: Deep features from DTransMIL are processed by an SVM classifier to simultaneously predict high-risk subtypes and recurrence risk. This integrated approach allows comprehensive characterization of PitNETs using only routine H&E-stained slides, demonstrating potential to support clinical decision-making through automated, multi-task analysis. WDL without distinct lineage, MLT male lactotroph tumours, SGST sparsely granulated somatotroph tumours, ASCT acidophil stem cell tumours, IPPT immature plurihormonal PIT1-lineage tumours, SCT silent corticotroph tumours, CCT Crooke cell tumours, NCT null cell tumours.

Dataset preprocessing

In order to cope with the challenges of large data scale, long processing time and high storage requirements faced by WSI in the task, and to achieve effective morphological feature extraction, this study designed and constructed a complete preprocessing process, which mainly covered several key steps: tissue region detection and image block segmentation, staining normalization and data enhancement. The specific implementation is as follows: First, the automatic segmentation algorithm in the CLAM toolkit⁴¹ is used to accurately distinguish the foreground (tissue area) and background of each WSI, automatically identify the effective tissue area and generate the corresponding tissue mask. On this basis, the tissue area is cropped into non-overlapping image blocks with a size of 1024 × 1024 pixels at ×10 magnification, and only the area containing the tissue structure is retained, so as to effectively filter out a large number of non-information background areas and reduce the calculation and storage costs of subsequent processing. Secondly, in order to reduce the inconsistency of staining style caused by the difference of scanning equipment, different batches of staining reagents and other factors, the Vahadane staining normalization method is applied to map all image blocks to a unified colour space, so as to eliminate the interference of technical variation on the extraction of pathological morphological features and ensure the consistency and comparability of data. Finally, during the model training phase, data augmentation was applied to the internal training data following the deep learning algorithm for colorectal cancer described in reference⁴². Specifically, operations such as random rotation, horizontal and vertical flipping were employed. This enhanced the diversity of the training data, improved the model’s robustness to variations in data distribution, and thereby strengthened its generalization performance.

Lineage prediction via macro–micro dual branches

In clinical practice, pathologists first examine slides at low magnification to identify key tissue structures and then zoom in on these areas for detailed cellular analysis. Inspired by this diagnostic process, we propose a novel PLIP-guided DTransMIL framework that integrates both macroscale and microscale features to achieve accurate lineage prediction.

The macrobranch is a transformer-based multiple instance learning (TransMIL) module that processes WSI data at a coarser scale. The macrobranch generates a macroattention map, highlighting high-attention regions in the image. Formally, given the extracted macroscale patch features ${{\{x}_i^{\mathrm{macro}}\}}_{i=1}^{N_{\mathit{macro}}}$ as input, the macrobranch extracts the corresponding patch embeddings $h_i^{\mathrm{macro}}=T_{\mathit{macro}}\left(x_i^{\mathrm{macro}}\right),$ where $T_{\mathit{macro}}(·)$ denotes the macroscale TransMIL feature extractor and $N_{\mathit{macro}}$ is the total number of macroscale patches. These patch-level features are then aggregated into a slide-level representation:

$$\begin{array}{c}{H}^{\mathrm{m}\mathrm{a}\mathrm{cro}}=\sum _{i=1}^{N_{\mathit{macro}}}{h}_i^{\mathrm{macro}}.\end{array}$$ 1

The attention weights ${{\{\alpha }_i^{\mathrm{macro}}\}}_{i=1}^{N_{\mathit{macro}}}$ are then computed as follows:

$$\begin{array}{c}{\alpha }_i^{\mathrm{macro}}=\mathit{softmax}\left(\frac{Q_s\left(x_i^{\mathrm{macro}}\right)\cdot {\left(K_s\left(x_i^{\mathrm{macro}}\right)\right)}^T}{\sqrt{d_k}}\right),\end{array}$$ 2

where $Q_s\left(·\right)$ and $K_s\left(·\right)$ are linear mappings of the input and $d_k$ is the scaling factor.

From this macroattention map ${{\{\alpha }_i^{\mathrm{macro}}\}}_{i=1}^{N_{\mathit{macro}}}$, we identify the K patches with the highest attention scores. These patches are considered the most representative for capturing the overall tissue characteristics. We then construct a binary mask map M as follows:

$$\begin{array}{c}M\left(i\right)=\left\{\begin{array}{cc}1& ,{\alpha }_i^{\mathrm{macro}}\mathit{is}\mathit{among}\mathit{the}\mathit{top}K\\ 0& ,\mathit{else}\end{array}\right).\end{array}$$ 3

The mask map $M$ serves as a guide to isolate these important patches. Using the established spatial correspondence between macro- and microscale patches, defined by a mapping function $\Phi \left(i\right)$, we select the corresponding microscale patches as follows:

$$\begin{array}{c}{S}^{\mathit{micro}}=\Phi \left(\{i,|,M\left(i\right)=1\}\right),\end{array}$$ 4

where $S^{\mathit{micro}}$ denotes the index set of the microscale patches selected on the basis of the macroscale attention map.

These selected patches, {$x_j^{\mathrm{micro}}$}, $j\in S^{\mathrm{m}\mathrm{i}\mathrm{cro}}$, are then fed into the Micro-TransMIL module, which operates at a finer spatial resolution. This module further analyses the patches, capturing more detailed and specific features related to the local cell composition and microstructural organization:

$$\begin{array}{c}{h}_j^{\mathrm{micro}}=T_{\mathit{micro}}\left(x_j^{\mathrm{micro}}\right),j\in S^{\mathrm{micro}},\end{array}$$ 5

where $T_{\mathit{micro}}\left(·\right)$ denotes the Micro-TransMIL feature extractor. Finally, the resulting microscale patch features are aggregated into a slide-level representation:

$$\begin{array}{c}{H}^{\mathrm{micro}}=\frac{1}{K}\sum _{j\in S^{\mathrm{m}\mathrm{i}\mathrm{cro}}}{h}_j^{\mathrm{micro}},j\in S^{\mathrm{m}\mathrm{i}\mathrm{cro}}.\end{array}$$ 6

After the unique macrobranch and the microbranch features are extracted from the tissue data, the derived features $H^{\mathrm{macro}}$ and $H^{\mathrm{micro}}$ are aggregated through concatenation. Furthermore, the combined features$H^{\mathit{joint}}$ are fed into a prediction module to distinguish the lineages of PitNETs. By integrating the information from both branches, the prediction model can consider both the coarse-level overall tissue architecture and the fine-level local cell features, thereby enhancing the accuracy and comprehensiveness of lineage prediction.

Prediction of high-risk subtypes and recurrence

Beyond lineage prediction, PitNETs demonstrate significant biological heterogeneity even within the same lineage, resulting in varied clinical outcomes. To further investigate the prognostic potential of the learned morphological representations, we utilize the deep features extracted during lineage classification to predict intralineage subtypes and recurrence outcomes.

Specifically, the concatenated slide-level features $H^{\mathit{joint}}$, derived from the PLIP-guided DTransMIL framework, encode both the macroarchitectural and the microcellular characteristics of each WSI. For subtype classification and recurrence prediction, we employ machine learning-based modules that aim to find optimal decision boundaries to separate different classes in the feature space. In our study, these features are input into support vector machines (SVMs) to separately estimate intralineage subtype probabilities and recurrence risk:

$$\begin{array}{c}{\hat{y}}_i^{\left(\mathit{subtype}\right)}=f_{\mathit{subtype}}\left(H_i^{\mathit{joint}}\right),{\hat{y}}_i^{\left(\mathit{re}c\right)}=f_{\mathit{rec}}\left(H_i^{\mathit{joint}}\right),\end{array}$$ 7

where ${\hat{y}}_i^{\left(\mathit{subtype}\right)}\in {\mathbb{R}}^C$ represents the predicted probability distribution across C subtypes within a lineage and ${\hat{y}}_i^{\left(\mathit{re}c\right)}\in [0,1]$ denotes the predicted recurrence probability.

In this way, the proposed framework not only enables accurate lineage classification but also extends to prognostic risk stratification, bridging the morphological phenotype with clinical outcomes of PitNETs.

Multiplex immunofluorescence staining and data analysis

To investigate the composition of different immune cells in PitNETs and enhance the interpretability of WSI-based recurrence prediction, we selected 22 tissue samples (11 primary and 11 matched recurrence cases) and performed multiplex immunofluorescence (mIF) analysis. These samples included 3 pairs of SF1 lineage cells, 3 pairs of TPIT lineage cells, 3 pairs of PIT1 lineage cells, and 2 pairs of PitNETs without distinct lineages. Two panels were constructed, which included the following markers: CD3, CD4, CD8, FoxP3, PD-1, Synaptophysin (Syn), CD20, CD68, CD11b, CD163, HLA-DR, and DAPI. The antibody clones used in this study are shown in Supplementary Table 4.

Multiplex immunofluorescence analysis was conducted using the FAIP-48T fully automated immunohistochemistry platform (Abcarta, USA). We employed a Tyramide Signal Amplification (TSA)-based 7-colour multiplex fluorescence immunohistochemistry kit (Abcarta, PS017) supplemented with the following components: TSA fluorophores (425 nm, 488 nm, 532 nm, 594 nm, 633 nm, and 680 nm emission wavelengths), DAPI nuclear counterstain, signal amplification buffer, HRP-conjugated goat anti-rabbit/mouse secondary antibody polymer, and endogenous peroxidase blocking reagent. All immunostained slides were digitally imaged using the HEIDSTAR high-throughput fluorescence scanning system (version 23.10.30). Quantitative and spatial analyses were performed using the HALO image analysis platform (version 3.6.4134.362), with particular emphasis on the HighPlex FL module for multiplex fluorescence quantification and spatial distribution assessment.

Spatial transcriptome analysis

After rigorous quality screening of the paraffin-embedded tissues, we selected 6 primary tumours and 6 matched recurrence tumours for spatial transcriptome sequencing, which included 2 pairs of the SF1 lineage, 3 pairs of the TPIT lineage, and 1 pair of PitNETs without a distinct lineage. Among the 32 recurrence cases, all three PIT1-lineage tumours were unable to undergo spatial transcriptomic sequencing due to inadequate FFPE block quality or insufficient tissue amount. One 5 µm section per TMA was applied to the Visium system for each FFPE slide (one TMA per 6.5 × 6.5 mm capture area) and stained with H&E. Tissues were prepared following the demonstrated protocols: Xenium In Situ for FFPE-Tissue Preparation Guide (CG000578) and Xenium In Situ for FFPE Tissue Deparaffinization & Decrosslinking (CG000580). Probe hybridization, washing, ligation, amplification, and cell segmentation staining were performed following a development version of the Xenium Prime In Situ Gene Expression with Optional Cell Segmentation Staining user guide. Postinstrument processing followed the demonstrated protocol: Xenium In Situ Gene Expression-Post-Xenium analyser H&E Staining (CG000613). The Xenium Prime 5 K Human Pan Tissue and Pathways panel was designed to enable comprehensive cell type and cell state identification using publicly available single-cell RNA sequencing data. The panel also covers canonical signalling pathways, as well as genes relevant to developmental biology and immuno-oncology and genes that are well known in the biomedical literature. The instrument run was performed following the Xenium Analyser User Guide (CG000584). The on-instrument analysis was run with a development version of Xenium Onboard Analysis v3.0.0.

RT-qPCR

Total RNA was extracted from 12 formalin-fixed paraffin-embedded (FFPE) clinical samples (6 primary and 6 recurrence) using the RNeasy FFPE Kit (Qiagen, Cat. No. 73504). RNA concentration and purity were assessed using a NanoDrop spectrophotometer. First-strand cDNA was synthesized from 500 ng of total RNA using a PrimeScript RT Reagent Kit (Takara). RT-qPCR was performed in triplicate on a QuantStudio 5 system with SYBR Green Master Mix (Applied Biosystems). Gene-specific primers were designed by Sangon Biotech (Shanghai) and validated for efficiency (90–110%). GAPDH was used as the internal reference gene. Relative gene expression was calculated using the 2^-△△Ct algorithm. The gene-specific primers used are listed in Supplementary Table 5.

Statistical analysis

The performance of the classification network for predicting PitNET lineages, high-risk subtypes, and recurrence was evaluated using receiver operating characteristic (ROC) curve analysis (with the area under the curve (AUC) as the primary metric). The optimal thresholds were determined by maximizing the Youden index while considering the accuracy, sensitivity, specificity, PPV, and F1 score. The calculation formulas are PPV = TP/(TP + FP), Sensitivity= TP/(TP + FN), Specificity= TN/(TN + FP), where TP (True Positive) refers to the number of samples that are actually of a target lineage and predicted as that lineage by the model, TN (True Negative) refers to the number of samples that are actually of non-target lineages and predicted as non-target lineages by the model, FP (False Positive) refers to the number of samples that are actually of non-target lineages but predicted as the target lineage by the model, and FN (False Negative) refers to the number of samples that are actually of a target lineage but predicted as non-target lineages by the model. Group differences were assessed using ANOVA with Tukey post hoc tests or Wilcoxon tests. Values with ∗: P < 0.05, ∗∗: P < 0.01, and ∗∗∗: P < 0.001 are indicative of statistical significance.

Data pre-processing, model development, and the generation of all statistical figures and tables were conducted using Python (version 3.9.18) and the deep learning platform PyTorch (version 1.7.7).

Author contributions

Design: Z.A.L., Z.F., W.M.H., W.H.B., X.J., W.D.Z.; Date: G.C.; W.W., X.J., T.X.H., C.L.Q., D.Z.G.; Manuscript revisions: L.A., W.M.H., W.H.B., X.J.

Data availability

The datasets generated and analyzed during the current study are not publicly available due to institutional data-sharing policies but are available from the corresponding author on reasonable request. Please e-mail the corresponding author, Haibo Wu, M.D., Ph.D., at wuhaibo@ustc.edu.cn.

Code availability

The underlying code for the procedures performed in this study can be accessed via this link: https://github.com/fz-ustc/DtransMIL.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-026-01444-8.