DeepPMD: A Comprehensive Deep Learning Framework for Primary-Metastatic Classification and Origin Prediction in Lung Adenocarcinoma – Multi-Center Whole Slide Image Validation

DeepPMD: A Comprehensive Deep Learning Framework for Primary-Metastatic Classification and Origin Prediction in Lung Adenocarcinoma – Multi-Center Whole Slide Image Validation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article DeepPMD: A Comprehensive Deep Learning Framework for Primary-Metastatic Classification and Origin Prediction in Lung Adenocarcinoma – Multi-Center Whole Slide Image Validation Li Liang, Guo Shujing, Jiabo MA, Li Jinbang, Li Danyi, Zhao Chenglong, and 16 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7819870/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Background Among solid malignancies, metastatic spread accounts for ~ 90% of cancer-related deaths. For unknown pulmonary lesions, accurate diagnosis is crucial for treatment. In diagnostically challenging cases like second primary tumor and cancer of unknown primary (CUP), traditional morphological analysis and immunohistochemistry (IHC) often remain inconclusive, delaying critical therapeutic intervention. Methods We developed a deep learning model (DeepPMD) that makes predictions using H&E stained whole slide images and basic clinical data. The model was trained and validated on a single-center cohort of 793 patients, then externally validated on three additional cohorts totaling 1187 patients from independent centers. Results DeepPMD achieved accurate prediction of tumor nature and site of origin, outperforming benchmarks in all tests. In external validation, the model achieved a macro-AUC of 0.974 (95% CI 0.949, 0.991) for origin prediction on excisional biopsies and 0.935 (95% CI 0.913, 0.953) on aspiration biopsies. Its diagnostic logic aligned with pathological criteria, and its CUP predictions showed high concordance with clinical predictions (consistency score 0.86). Conclusion DeepPMD is an accurate, generalizable, and interpretable AI tool for identifying pulmonary tumor origins from WSIs. It can provide probability-ranked potential primary sites to guide targeted IHC testing, optimizing diagnostic workflow, shortening turnaround times, and conserving tissue samples. Biological sciences/Cancer/Lung cancer/Small-cell lung cancer Biological sciences/Computational biology and bioinformatics/Software Tumor Origin Prediction Computational Pathology Cancer of Unknown Primary (CUP) Deep learning Whole Slide Images Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Lungs are among the most common sites for malignant tumor metastasis, with an estimated 20% to 54% of patients with extrathoracic malignancies developing pulmonary metastases during their disease course, resulting in a dramatic decline in prognosis and survival rates. 1 For instance, the 5-year survival rate for patients with localized colorectal cancer is 91%, a figure that plummets to 14% following distant metastasis. 2 Therefore, for any pulmonary lesion, accurately identifying the origin of a tumor is a critical prerequisite for guiding subsequent therapy and improving patient outcomes. 3 – 6 Immunohistochemistry (IHC), valued for its efficiency, accuracy, and affordability, is the primary method for identifying tumor origins. However, its clinical application is complicated by the need not only to exclude second primary lung cancers, which can be morphologically indistinguishable from metastases, 7 but also to address the diagnostic challenge posed by Cancer of Unknown Primary (CUP) . 8–12 Pathologists, particularly those with limited experience in primary care settings, tend to expand IHC panels to avoid misdiagnosis, which increases medical costs, sample consumption, and patient burden. The digitization of pathology through the Whole Slide Images (WSIs) has recently created unprecedented opportunities to apply artificial intelligence (AI). Deep learning models have demonstrated expert-level performance in a range of histopathological tasks, including Gleason grading of prostate cancer biopsies 13 – 15 and predicting survival in colorectal cancer. 16 – 18 This success has extended robustly into pulmonary pathology, where models can now predict genetic mutations, 19,20 patient survival, 21 immunotherapy response, 22 and even future metastasis risk directly from H&E slides. 23 Beyond these applications lies the fundamental diagnostic challenge of identifying a tumor's origin, particularly for Cancer of Unknown Primary (CUP). The landmark TOAD model, 24 firstly established the feasibility of using AI to address this challenge by predicting a primary site from histopathology slides. However, critical limitations remain in the current AI approaches. First, a lack of validation on multi-center, multi-device data casts doubt on their generalizability, as algorithmic performance may not translate across heterogeneous data from different institutions. Second, most models rely solely on morphological information, ignoring crucial clinical data, further their decision-making processes often lack the transparency needed for clinical trust. To address these limitations, we aimed to develop and validate DeepPMD (Deep Learning for Primary-Metastatic Differentiation), a deep learning framework that integrates multi-center WSI data with clinical information to accurately differentiate primary from metastatic tumors in the lung and identify the origin of tumors. We further sought to ensure model transparency through interpretability analysis and rigorously validate its performance in a large, multi-center cohort to assess its real-world clinical utility. Methods Study design and participants In this national, multi-center and retrospective study, we established DeepPMD to effectively identify primary and metastatic tumors and predict the primary tumor site. To develop and internally validate DeepPMD, we established a discovery cohort of 773 patients (providing 1314 WSIs) from Southern Medical University (SMU) between January 2009 and March 2023 (Figure S1). Then we collected three independent cohorts from three cities throughout China for external validation: External #1 was from the First Affiliated Hospital of Shandong First Medical University, comprising 453 patients (453 WSIs) between 2004 and 2014; External #2 was from The Fourth Hospital of Hebei Medical University, with 547 patients (1007 WSIs) between 2019 and 2024; and External #3 was from The Sixth Affiliated Hospital of Guangzhou Medical University, with 187 patients (318 WSIs) between 2010 and 2023. In total, the external validation set consisted of 1187 patients and 1778 WSIs (Table S1). All WSIs were scanned at 20× and stored in SVS, TIFF, and MRXS formats and finally converted to SVS format for training. To ensure robust model training and evaluation, pathological types that were present across all four cohorts but had fewer than 2 cases in any single cohort were consolidated into an "Others" category. This resulted in a final classification schema of six categories: lung, gastrointestinal, breast, kidney, liver, and others. Inclusion and exclusion criteria and further details are provided in the appendix (Supplementary Criteria, Supplementary Figure S1). This study was approved by the Ethics Committee and Institutional Review Board of Nanfang Hospital, Southern Medical University (Approval No. NFEC-2024-379). All experiments were conducted in accordance with the Declaration of Helsinki. Since the study did not involve the collection of private information, we requested a waiver of informed consent for the internal, retrospectively collected data, which was granted by all four participating centres. Model Development The model was developed and internally validated on 773 patients from the Southern Medical University (SMU) cohort, which were divided at the patient level into a training set (n=541), a test set (n=155), and an internal validation set (n=77) at a 7:2:1 ratio. We selected the H-optimus-1 model, which ranked first in overall performance on lung cancer-related diagnostic tasks and excelled in primary site prediction according to a large-scale benchmark study for pathology foundation models (PathBench) as our foundation model. 25 H-optimus-1 is a vision foundation model with a ViT-G/14 architecture, pre-trained on a dataset of approximately one million WSIs, covering more than 50 tissue types and three scanner devices. To enhance classification performance, we followed the approach of the TOAD model by integrating these patch-level features with patient-level clinicopathological data (gender and age) to create multimodal inputs. Our architecture incorporates two parallel processing branches with gated attention mechanisms: one for primary versus metastatic cancer classification and another for the six-class primary site prediction. This multi-task attention pooling dynamically weights patch features according to their diagnostic relevance for each specific task. These weighted features are aggregated into a comprehensive slide-level representation and combined with the clinical information for the final prediction. Further details on WSI preprocessing and model architecture are available in the Supplementary Methods. Model Validation DeepPMD was evaluated through a multi-stage validation process to assess its accuracy, generalizability, and robustness. The model was firstly validated on the internal validation set, followed by testing on three independent external cohorts to assess real-world generalizability. Furthermore, its robustness was assessed on different specimen types (excisional and aspiration biopsies) within the whole external datasets, and a per-class performance analysis was conducted for each primary origin. Then we validated the clinical relevance and decision-making process of the model via a multi-step interpretability analysis. Firstly, t-SNE was used to visualize the feature space of the learned features from both DeepPMD and TOAD to assess class separability. Next, to evaluate DeepPMD's attention mechanism, a senior pathologist scored attention heatmaps from 100 randomly selected cases from each of the three external validation cohorts. The scoring was based on the degree of overlap between high-attention regions and actual tumor areas, using a 5-point scale: 1 (0–20%), 2 (20–40%), 3 (40–60%), 4 (60–80%), and 5 (80–100%). Finally, to understand the morphological basis of the model's predictions, high-attention patches from the heatmaps were clinically interpreted by the pathologist based on the pathological criteria for differentiating primary lung adenocarcinoma from metastatic carcinomas. This heatmap scoring and patch-level interpretation were performed only for the DeepPMD model. The performance of DeepPMD was benchmarked against the state-of-the-art TOAD model across all evaluations. Statistical Analysis Model performance was evaluated for both the binary primary-metastatic classification and the multi-class origin prediction tasks. The overall discriminative ability of the models was assessed using the area under the receiver operating characteristic curve (AUC). We also reported accuracy (ACC) and the F1 score, which considers both precision and recall, to provide a comprehensive evaluation. To account for class imbalance in the multi-class task, we reported macro-averaged (the unweighted mean of per-class scores) and weighted-averaged (weighted by per-class support) F1 scores. The stability of all performance metrics was quantified by 95% confidence intervals (CIs), generated using non-parametric bootstrapping with 1,000 resamplings. To compare the performance between the DeepPMD and TOAD models, the two-sided Wilcoxon signed-rank test was used, with a p-value of less than 0.05 considered statistically significant. Results In the SMU cohort, 1347 WSIs were available from 773 individuals (427 [53.85%] men and 366 [46.015%] women) (Fig. 1 B). Median age was 58 years [IQR 18–86]. For the external validation cohorts, we enrolled 1778 WSIs from 1187 patients across three medical centers around from north to south in China (701 [59.06%] men and 486 [40.94%] women; median age 61 years [IQR 19–86]). The cohorts exhibited significant differences in biopsy types and technical parameters, with aspiration biopsies being predominant in the internal cohort (45.67%), while excisional biopsies were more common in the external cohort (59.06%). Additionally, different brands of WSI scanners (Leica, Ibingli, and 3DHISTECH) were used, ensuring the diversity of the dataset for a more comprehensive evaluation of the model's generalization capability. For the fundamental performance of the model, DeepPMD accurately differentiated a primary tumor from a metastatic tumor and identifying the origin of tumor (Macro-AUC 0.993 [95%CI 0.984, 0.999]), and identified the origin of each metastatic tumor (Macro-AUC 0.992 [95%CI 0.985, 0.998]) in the whole internal validation set (Fig. 2 A). On each external cohort, the value of Macro-AUC was from 0.850 (95%CI 0.813, 0.884) to 0.981 (95%CI 0.962, 0.995) for differentiation between primary and metastatic cancer, and from 0.914 (95% CI 0.889, 0.935) to 0.981(95% CI 0.823, 0.996) for primary tumor site prediction (Fig. 2 B). And on the pooled external validation cohort, DeepPMD achieved stable and excellent performance regardless of specimen types (Fig. 3 A-B). On excisional biopsy samples, which typically provide more comprehensive tissue architecture, DeepPMD demonstrated exceptional performance. It had a Macro-AUC value of 0.962 (95%CI 0.950, 0.973), Macro-ACC value of 0.889 (95%CI 0.869, 0.910) and Macro-F1 value of 0.837 (95%CI 0.808, 0.865) for discrimination of primary versus metastatic, while a Macro-AUC value of 0.974 (95%CI 0.949, 0.991), Macro-ACC value of 0.850 (95%CI 0.747, 0.925) and Macro-F1 value of 0.767 (95% CI 0.659, 0.827) for primary tumor site prediction. In the more challenging aspiration biopsies, the model had a Macro-AUC value of 0.944 (95%CI 0.920, 0.966), Macro-ACC value of 0.844 (95%CI 0.801, 0.883) and Macro-F1 value of 0.841 (95%CI 0.799, 0.877) for discrimination of primary versus metastatic, while a Macro-AUC value of 0.935 (95%CI 0.913, 0.953), Macro-ACC value of 0.728 (95%CI 0.670, 0.781) and Macro-F1 value of 0.756 (95% CI 0.702, 0.801) for primary tumor site prediction. Meanwhile, recognizing the imperative for high reliability in clinical applications and the potential for macro-averaging to obscure poor performance in minority classes, we conducted a stratified analysis of the model's primary site prediction performance on aspiration versus excisional biopsy samples from all the external cohorts (Fig. 3 C). On aspiration biopsy samples, the DeepPMD model demonstrated excellent performance across all cancer types, with particularly strong results in predicting gastric & colorectal cancer (AUC = 0.965, 95% CI 0.947–0.979) and renal cancer (AUC = 0.966, 95% CI 0.924–0.992). The model similarly exhibited a high level of accuracy on excisional biopsy samples, achieving remarkable discrimination for renal cancer with an AUC of 0.999 (95% CI 0.996–1.000). In a combined analysis of both biopsy types, the DeepPMD model maintained its robust predictive capabilities across all primary sites, further demonstrating its generalizability and reliability, irrespective of the biopsy acquisition method. When benchmarked against the TOAD model, DeepPMD demonstrated significantly higher performance (p < 0.001 for all key comparisons). On the internal validation set, DeepPMD's Macro-AUC advantage over TOAD was modest (10.8% for the binary task and 12.1% for the six-class task). However, this gap widened substantially on external cohorts, with the margin for the binary task on External #1 increasing to 31.2% (0.850 vs 0.648). TOAD's poor generalization was most pronounced in the six-class prediction task, where its Macro-ACC only dropped to 0.269 on External #1, and on aspiration biopsies, where its Macro-AUC (0.731) was markedly lower than DeepPMD's (0.935) . This performance advantage stems from DeepPMD’s ability to learn more discriminative features, as visualized by t-SNE dimensionality reduction (Fig. 4 , Supplementary Figure S2). The DeepPMD model formed well-demarcated feature clusters for both binary (primary vs. metastatic) and six-class (primary site prediction) tasks on the combined dataset. This clear clustering was consistently reproduced in the internal and three independent external validation sets, confirming the robust cross-center generalizability of the learned features. In the t-SNE visualization of the combined dataset, tumors from different origins formed dense and distinct clusters. Notably, the clusters for lung adenocarcinoma (LUAD) and breast cancer exhibited some overlap, which is primarily attributable to their shared histopathological similarities, such as glandular architectures, abundant eosinophilic cytoplasm, and comparable nuclear atypia. This observation was corroborated by the model's confusion matrix: across all samples, 56 LUAD cases were misclassified as breast cancer, and 21 breast cancer cases were misclassified as LUAD (Supplementary Figure S3). This demonstrates that DeepPMD's confusion is not random error but reflects a learned morphological discrimination logic analogous to that of human pathologists. In contrast, the feature space of the TOAD model appeared disorganized. Particularly in the external validation sets, feature points from different classes were severely intermixed, lacking discernible cluster structures and clear boundaries. This evidence strongly indicates that DeepPMD learns efficient and robust discriminative features that successfully generalize to heterogeneous, multi-center data, whereas the TOAD model fails to learn effective universal features. To evaluate the model's interpretability, attention heatmaps from all external validation cohorts were systematically scored by a senior pathologist (Fig. 5 B). The results showed that the proportions of cases with high-focus scores (4 or 5) were 86.28% in External #1, 90.30% in External #2, and 89.84% in External #3. This demonstrates the model's consistent and accurate localization ability across datasets from different medical centers. Then we analyzed its high-attention regions (Fig. 5 C). Analysis of high-attention patches confirmed the model's decisions were based on key histomorphological criteria: for primary lung adenocarcinoma, the model focused on areas with mixed growth patterns, particularly the highly specific lepidic growth pattern; for metastatic cancers, its attention was origin-specific, identifying the classic triad of serrated glands, nuclear palisading, and “dirty” necrosis in colorectal metastases; locating the single-file line (Indian file) arrangements characteristic of lobular breast carcinoma; concentrating on thick trabecular structures of eosinophilic polygonal cells separated by a rich sinusoidal network in hepatocellular carcinoma; and capturing nests of clear cells within a dense, thin-walled capillary network in clear cell renal carcinoma. This visual evidence strongly suggests that the model has learned to recognize the tissue-specific morphological cues that are critical for differential diagnosis, rather than relying on artifacts or spurious features, thereby confirming the pathological basis of its decisions. Finally, to assess the model's potential utility in addressing the clinical challenge of Cancer of Unknown Primary (CUP), we designed a consistency analysis(Fig. 5 A). In this analysis, the model's top-k predictions were compared against the k differential diagnoses provided clinically; an overlap of over 50% was considered consistent. The analysis revealed that DeepPMD's predictions achieved a consistency score of 0.86 with the clinical diagnostic path, whereas the TOAD model scored only 0.37. Based on this performance, we developed SmartPath, a system that generates a probability-ranked list of potential primary sites for pathologists directly from WSIs (Supplementary Figure S4). This approach optimizes the diagnostic workflow by guiding the selection of targeted IHC panels, thereby shortening diagnostic turnaround time, increasing precision, and conserving tissue, particularly in cases with limited clinical data or tissue samples Discussion In this study, we developed and validated DeepPMD, a deep learning model that accurately differentiates between primary and metastatic lung tumors and predicts their primary origin using only standard WSIs and basic clinical data. Evaluated on a large-scale cohort from four independent medical centers using three different scanners, DeepPMD demonstrated exceptional accuracy and robust generalizability, significantly outperforming existing benchmark models. Furthermore, our interpretability analysis revealed that the model's diagnostic logic is highly concordant with the histopathological criteria used by pathologists, thereby reinforcing its reliability for clinical decision-making. In clinical practice, when faced with a pulmonary tumor of unknown origin, pathologists rely on patient history and histomorphological features to select a broad IHC panel to ascertain its primary or metastatic nature and potential origin. In complex cases or with less experienced pathologists, a more extensive IHC panel is often chosen, which prolongs diagnostic turnaround time, consumes limited biopsy tissue, and increases healthcare costs. DeepPMD is poised to optimize this diagnostic workflow. By inputting a WSI into the system interface, pathologists can instantaneously receive a probability-ranked list of potential primary sites, enabling the selection of a more targeted, narrow-spectrum IHC panel. This approach can significantly shorten the diagnostic timeline, enhance precision, and conserve valuable time for therapeutic planning, proving especially beneficial in cases with limited clinical history or scant tissue samples. Moreover, as conventional broad-panel IHC can sometimes yield inconclusive results, leading to treatment delays and potential disputes, our model provides an objective, data-driven reference that reduces diagnostic uncertainty, supports the final pathological conclusion, and may mitigate risks associated with diagnostic ambiguity. The robustness of the model is fundamental to its clinical applicability. By validating DeepPMD across four centers with different WSI scanners and clinical practices, we have shown that the morphological features it learns are robust and generalizable. Furthermore, our interpretability analyses confirmed that the model's diagnostic logic aligns closely with the visual criteria used by pathologists, enhancing its trustworthiness for clinical decision-making. Despite these encouraging results, our study has several limitations. First, the current model was trained and validated on a selection of the most common metastatic tumor types, and its performance on rarer cancer origins has yet to be explored. Second, this was a retrospective study. Although rigorously validated on multi-center data, a prospective study has not yet been conducted to evaluate its real-world impact on clinical workflows, such as direct effects on treatment decisions. Future work will focus on expanding the range of tumor types and validating the clinical utility of DeepPMD in a prospective trial. In conclusion, DeepPMD demonstrates immense potential as an accurate, efficient, and interpretable AI-powered tool for the challenging task of tumor origin identification. By optimizing IHC testing strategies, it promises to shift the paradigm in pathological diagnosis, ultimately benefiting patients. Declarations Contributors S.G.: Conceptualization, Data Curation, Validation, Methodology, Visualization, Writing – Original Draft, Writing – Review & Editing. J.M.: Methodology, Software, Formal Analysis,Validation, Visualization. J.L.: Conceptualization, Methodology, Data Curation, Validation, Investigation. C.Z.: Validation, Investigation. D.L., X.W.: Writing – Review & Editing. F.Z., Y.X., Z.G.: Visualization. Z.Z., Y.T., A.A., W.Q., X.Z., B.L., Q.X., Z.W.: Data Curation. Y.W., Y.L., J.C., H.C., L.L.: Supervision, Project Administration, Funding Acquisition. L.L., Y.W., Y.L., J.C. also provided: Resources. H.C. also contributed to: Software. All authors reviewed and approved the final manuscript. Data availability Datasets are not publicly available due to patient privacy considerations. Code availability The underlying code for this study is publicly available in the GitHub repository (https://github.com/birkhoffkiki/DeepPMD). Declaration of interests We declare no competing interests. Acknowledgments This work was supported by grants from the National Natural Science Foundation of China (82273358), Chongqing Technology innovation and application development special major project (CSTB2024TIAD-STX0003). References Budczies J, von Winterfeld M, Klauschen F, et al. The landscape of metastatic progression patterns across major human cancers. Oncotarget. 2015;6(1):570-83. Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin. 2024;74(1):12-49. doi:10.3322/caac.21820 Chou TY, Dacic S, Wistuba I, et al. Differentiating separate primary lung adenocarcinomas from intrapulmonary metastases with emphasis on pathological and molecular considerations: recommendations from the International Association for the Study of Lung Cancer Pathology Committee. J Thorac Oncol. 2025;20(3):311-30. doi:10.1016/j.jtho.2024.11.016 Chen K, Liu A, Wang C, et al. Multidisciplinary expert consensus on diagnosis and treatment of multiple lung cancers. Med. 2025;6(4):100643. doi:10.1016/j.medj.2025.100643 Brims F, McWilliams A, Williamson J, Siemienowicz M, Leong TL; Thoracic Society of Australia & New Zealand Lung Cancer Working Party. The TSANZ Practical Guide for Clinicians in the Management of Screen- and Incidentally-Detected Nodules. Respirology. 2025;30(7):558-73. doi:10.1111/resp.70065 Anagnostopoulos AK, Gaitanis A, Gkiozos I, et al. Radiomics/radiogenomics in lung cancer: basic principles and initial clinical results. Cancers (Basel). 2022;14(7):1657. doi:10.3390/cancers14071657 Chuang SC, Scélo G, Lee YC, et al. Risks of second primary cancer among patients with major histological types of lung cancers in both men and women. Br J Cancer. 2010;102(7):1190-5. doi:10.1038/sj.bjc.6605616 Kato S, Alsafar A, Walavalkar V, Hainsworth J, Kurzrock R. Cancer of unknown primary in the molecular era. Trends Cancer. 2021;7(5):465-77. doi:10.1016/j.trecan.2020.11.002 Krämer A, Bochtler T, Pauli C, et al. Cancer of unknown primary: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol. 2023;34(3):228-46. doi:10.1016/j.annonc.2022.11.013 Economopoulou P, Mountzios G, Pavlidis N, Pentheroudakis G. Cancer of unknown primary origin in the genomic era: elucidating the dark box of cancer. Cancer Treat Rev. 2015;41(7):598-604. doi:10.1016/j.ctrv.2015.05.010 Boscolo Bielo L, Belli C, Crimini E, et al. Cancers of unknown primary origin: real-world clinical outcomes and genomic analysis at the European Institute of Oncology. Oncologist. 2024;29(6):504-10. doi:10.1093/oncolo/oyae038 Spurgeon L, Mitchell C, Cook N, Conway AM. Cancer of unknown primary: the hunt for its elusive tissue-of-origin - is it time to call off the search?. Br J Cancer. Published online July 4, 2025. doi:10.1038/s41416-025-03073-7 Nagpal K, Foote D, Tan F, et al. Development and validation of a deep learning algorithm for Gleason grading of prostate cancer from biopsy specimens. JAMA Oncol. 2020;6(9):1372-80. doi:10.1001/jamaoncol.2020.2485 Bulten W, Kartasalo K, Chen PC, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nat Med. 2022;28(1):154-63. doi:10.1038/s41591-021-01620-2 Campanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med. 2019;25(8):1301-9. doi:10.1038/s41591-019-0508-1 Mahajan A, Kania V, Agarwal U, et al. Deep-learning-based predictive imaging biomarker model for EGFR mutation status in non-small cell lung cancer from CT imaging. Cancers (Basel). 2024;16(6):1130. doi:10.3390/cancers16061130 Skrede OJ, De Raedt S, Kleppe A, et al. Deep learning for prediction of colorectal cancer outcome: a discovery and validation study. Lancet. 2020;395(10221):350-60. doi:10.1016/S0140-6736(19)32998-8 Foersch S, Glasner C, Woerl AC, et al. Multistain deep learning for prediction of prognosis and therapy response in colorectal cancer. Nat Med. 2023;29(2):430-9. doi:10.1038/s41591-022-02134-1 Coudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559-67. doi:10.1038/s41591-018-0177-5 Babu B, Singh J, Salazar González JF, et al. A narrative review on the role of artificial intelligence (AI) in colorectal cancer management. Cureus. 2025;17(2):e79570. doi:10.7759/cureus.79570 Pan X, Lin H, Han C, et al. Computerized tumor-infiltrating lymphocytes density score predicts survival of patients with resectable lung adenocarcinoma. iScience. 2022;25(12):105605. doi:10.1016/j.isci.2022.105605 Rakaee M, Tafavvoghi M, Ricciuti B, et al. Deep learning model for predicting immunotherapy response in advanced non-small cell lung cancer. JAMA Oncol. 2025;11(2):109-18. doi:10.1001/jamaoncol.2024.5356 Zhou H, Watson M, Bernadt CT, et al. AI-guided histopathology predicts brain metastasis in lung cancer patients. J Pathol. 2024;263(1):89-98. doi:10.1002/path.6263 Lu MY, Chen TY, Williamson DFK, et al. AI-based pathology predicts origins for cancers of unknown primary. Nature. 2021;594(7861):106-10. doi:10.1038/s41586-021-03512-4 Ma J, Xu Y, Zhou F, et al. PathBench: a comprehensive comparison benchmark for pathology foundation models towards precision oncology. arXiv. 2025 May 26. doi:10.48550/arXiv.2505.20202 Additional Declarations There is NO Competing Interest. Supplementary Files supplementarymaterial2.0.pdf supplementary material Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7819870","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":532216640,"identity":"df0cef1d-775c-4f6c-9c7f-fdcc2c091698","order_by":0,"name":"Li Liang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAlElEQVRIiWNgGAWjYBACPnbGxgcQZgKRWtiYGZsNSNXCwCZBqhbmtsqfOYcZ+NlzDBh+7iDOYW23ebcdZpDseWPA2HuGWC2MQC0GN3IMgGwitRT+BGqxJ0kLA8hhBhIkaGmW5t2WziNx5lnBwV5itPCztz/8+HObtRx/e/LGBz+J0QIDPCDiAAkaRsEoGAWjYBTgAwCBISydTrd14wAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0001-5302-2754","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":true,"prefix":"","firstName":"Li","middleName":"","lastName":"Liang","suffix":""},{"id":532216641,"identity":"e6c598f1-0c27-4a80-bfe6-6d11cbb2371e","order_by":1,"name":"Guo Shujing","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Guo","middleName":"","lastName":"Shujing","suffix":""},{"id":532216642,"identity":"4c02a5c4-dfd0-493b-aad9-f3bfbb7c06ae","order_by":2,"name":"Jiabo MA","email":"","orcid":"https://orcid.org/0000-0001-8532-4466","institution":"Hong Kong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Jiabo","middleName":"","lastName":"MA","suffix":""},{"id":532216643,"identity":"534ec59e-c08e-4fc3-b50e-189be0cab0d4","order_by":3,"name":"Li Jinbang","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Jinbang","suffix":""},{"id":532216644,"identity":"93489d2d-2770-49b7-a079-0823e5fa6749","order_by":4,"name":"Li Danyi","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Danyi","suffix":""},{"id":532216645,"identity":"7717fbca-2b93-4f26-a591-8a6c43222ae5","order_by":5,"name":"Zhao Chenglong","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zhao","middleName":"","lastName":"Chenglong","suffix":""},{"id":532216646,"identity":"ca494a2f-35f8-4ba9-8625-df5a29a1803a","order_by":6,"name":"Zhengyu ZHANG","email":"","orcid":"","institution":"Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zhengyu","middleName":"","lastName":"ZHANG","suffix":""},{"id":532216647,"identity":"976970e3-46c4-4142-a49d-566c0c0d6128","order_by":7,"name":"Ying Tan","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Tan","suffix":""},{"id":532216648,"identity":"e72337e9-8f51-471a-945e-a3e85bbf8766","order_by":8,"name":"Fengtao Zhou","email":"","orcid":"","institution":"The Hong Kong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Fengtao","middleName":"","lastName":"Zhou","suffix":""},{"id":532216649,"identity":"b127f7c9-3861-4bf1-b70d-06b62f31c330","order_by":9,"name":"Yingxue XU","email":"","orcid":"https://orcid.org/0000-0002-9657-3107","institution":"The Hong Kong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Yingxue","middleName":"","lastName":"XU","suffix":""},{"id":532216650,"identity":"40986b1b-007b-4814-9a28-e572a5044c97","order_by":10,"name":"Zhengrui GUO","email":"","orcid":"","institution":"The Hong Kong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Zhengrui","middleName":"","lastName":"GUO","suffix":""},{"id":532216651,"identity":"8437bb38-eb21-4043-842c-05d0f59b42f1","order_by":11,"name":"Xi Wang","email":"","orcid":"","institution":"The Hong Kong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Xi","middleName":"","lastName":"Wang","suffix":""},{"id":532216652,"identity":"03cdd792-0ef0-4730-95af-807c2afc5b14","order_by":12,"name":"Aimaier Aihetaimu","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Aimaier","middleName":"","lastName":"Aihetaimu","suffix":""},{"id":532216653,"identity":"ab91c627-fbfd-4ad7-91f4-cb6335775e8f","order_by":13,"name":"weihao qiu","email":"","orcid":"","institution":"Nanfang Hospital, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"weihao","middleName":"","lastName":"qiu","suffix":""},{"id":532216654,"identity":"0c08475f-20fa-45fe-b97d-fb7553c77fbb","order_by":14,"name":"Xiaohui Zhu","email":"","orcid":"","institution":"Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xiaohui","middleName":"","lastName":"Zhu","suffix":""},{"id":532216655,"identity":"7fe68812-5e8f-4864-8288-6c2f12b69492","order_by":15,"name":"Li Bingbing","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Bingbing","suffix":""},{"id":532216656,"identity":"7ffaf67d-21cc-478d-9966-80907a041666","order_by":16,"name":"Wang Zhen","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Wang","middleName":"","lastName":"Zhen","suffix":""},{"id":532216657,"identity":"c50a8788-d029-4ed3-baab-555f4273408d","order_by":17,"name":"Xie Qi","email":"","orcid":"","institution":"Department of Pathology, School of Basic Medical Sciences, Southern Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xie","middleName":"","lastName":"Qi","suffix":""},{"id":532216658,"identity":"f1f41e86-8e2c-41b0-8772-c862edd8bde2","order_by":18,"name":"Wang Yanfen","email":"","orcid":"","institution":"Department of Pathology, The Affiliated Qingyuan Hospital (Qingyuan People's Hospital), Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Wang","middleName":"","lastName":"Yanfen","suffix":""},{"id":532216659,"identity":"39ab0f33-df31-471d-be15-85b2b53eaf56","order_by":19,"name":"Yueping Liu","email":"","orcid":"https://orcid.org/0000-0002-4582-114X","institution":"The Forth Hospital of Hebei Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yueping","middleName":"","lastName":"Liu","suffix":""},{"id":532216660,"identity":"15e71683-1b6b-43cb-9957-34ac00d344d2","order_by":20,"name":"Cui Jing","email":"","orcid":"","institution":"Department of Pathology, The First Affiliated Hospital of Shandong First Medical University \u0026 Shandong Provincial Qianfoshan Hospital","correspondingAuthor":false,"prefix":"","firstName":"Cui","middleName":"","lastName":"Jing","suffix":""},{"id":532216661,"identity":"70476fff-031c-4e97-afdc-6dcf0461232e","order_by":21,"name":"Hao Chen","email":"","orcid":"https://orcid.org/0000-0002-8400-3780","institution":"The Hong Kong University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Hao","middleName":"","lastName":"Chen","suffix":""}],"badges":[],"createdAt":"2025-10-09 16:55:53","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7819870/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7819870/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95104195,"identity":"597a274c-8be0-4021-b2b0-48e39c9be035","added_by":"auto","created_at":"2025-11-04 10:27:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":776093,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe framework of the DeepPMD model and clinicopathological characteristics of the study cohorts.\u003c/strong\u003e(A) Schematic diagram of the DeepPMD workflow. The process begins with the acquisition of excisional or aspiration biopsy samples, followed by digitization into whole-slide images (WSIs). The model then processes these WSIs alongside basic clinical information, utilizing a gated attention-based multiple instance learning architecture to generate predictions. (B) Distribution of clinicopathological and demographic characteristics of the internal cohort and three external validation cohorts (#1, #2, and #3). The charts display the distribution of primary tumor diagnoses, patient age, patient sex, and biopsy sample types for each cohort.\u003c/p\u003e","description":"","filename":"Manuscript6.011.png","url":"https://assets-eu.researchsquare.com/files/rs-7819870/v1/33e861403c45213172879678.png"},{"id":95104197,"identity":"c103d020-5cd5-45e6-ae37-c593d9755815","added_by":"auto","created_at":"2025-11-04 10:27:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":26963,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance of DeepPMD and TOAD Models Across All Cohorts. \u003c/strong\u003eComparative performance of the DeepPMD and TOAD models evaluated on the internal cohort, three independent external cohorts (#1, #2, #3), and a combined external cohort. The bar charts display the Macro-AUC, with error bars representing 95% confidence intervals. (A) Model performance on the binary classification task of distinguishing primary from metastatic tumors. (B) Model performance on the six-class primary site prediction task. Across all datasets and for both tasks, DeepPMD consistently outperforms the TOAD benchmark model.\u003c/p\u003e","description":"","filename":"Manuscript6.012.png","url":"https://assets-eu.researchsquare.com/files/rs-7819870/v1/858bdede0778c15033d811e5.png"},{"id":95225401,"identity":"f634d5df-5f61-4002-8507-3acd2e7ecdcd","added_by":"auto","created_at":"2025-11-05 16:24:59","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":73922,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance of DeepPMD and TOAD Stratified by Biopsy Type and Primary Tumor Origin. \u003c/strong\u003e(A, B) Bar charts comparing the Macro-AUC of the DeepPMD and TOAD models on data stratified by biopsy type (excisional, aspiration, and all combined).\u003cstrong\u003e \u003c/strong\u003ePanel (A) shows performance on the binary classification task (primary vs. metastatic), while panel (B) shows performance on the six-class primary site prediction task. Error bars represent 95% confidence intervals. (C) Receiver Operating Characteristic (ROC) curves for the six-class primary site prediction task, providing a detailed per-class performance comparison. For each primary tumor type, the curves illustrate the model performance on excisional (EXC), aspiration (ASP), and all combined (All) samples. The corresponding Area Under the Curve (AUC) values are listed in the legends.\u003c/p\u003e","description":"","filename":"Manuscript6.013.png","url":"https://assets-eu.researchsquare.com/files/rs-7819870/v1/b69cb0dda7e724cb1c9dbc25.png"},{"id":95104194,"identity":"ca08ff88-59e6-4fe4-b72b-d908a35b1c18","added_by":"auto","created_at":"2025-11-04 10:27:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":110070,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003et-SNE visualization of feature representations learned by the DeepPMD and TOAD models. \u003c/strong\u003eThis figure uses t-SNE (t-distributed Stochastic Neighbor Embedding) to visualize the feature spaces learned by the models on all the datasets. The clustering of data points demonstrates the separability of features for different diagnostic tasks. (A): Feature distribution of DeepPMD model in binary classification task of primary and metastatic lung adenocarcinoma. (B): Feature distribution of DeepPMD model in six-class classification task of predicting tumor primary sites. (C): Feature distribution of TOAD model in binary classification task of primary and metastatic lung adenocarcinoma. D: Feature distribution of DeepPMD model in six-class classification task of predicting tumor primary sites.\u003c/p\u003e","description":"","filename":"Manuscript6.014.png","url":"https://assets-eu.researchsquare.com/files/rs-7819870/v1/716a74a2f1116d8c48ec73da.png"},{"id":95104196,"identity":"58275a8a-6790-4b2d-9ff3-70edf5e847d1","added_by":"auto","created_at":"2025-11-04 10:27:07","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1291137,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eInterpretability and Clinicopathological Correlation of the DeepPMD Model. \u003c/strong\u003e(A) Consistency analysis between model predictions and clinical differential diagnoses for Cancer of Unknown Primary (CUP). (B) Pathologist scoring of the overlap between DeepPMD's high-attention regions and actual tumor areas on a 5-point scale (1=0-20% overlap, 5=80-100% overlap). (C) Representative high-attention patches showing the histomorphological features identified by DeepPMD for each primary tumor class.\u003c/p\u003e","description":"","filename":"Manuscript6.015.png","url":"https://assets-eu.researchsquare.com/files/rs-7819870/v1/5671d2b30ab2a5d98558da83.png"},{"id":95230425,"identity":"1ec4c81e-7829-44fb-83ea-0e549ae6361d","added_by":"auto","created_at":"2025-11-05 16:37:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2902767,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7819870/v1/5a5a6f4a-7120-4ec5-a7e4-afa591fef545.pdf"},{"id":95104199,"identity":"fee6ce7b-12e1-463f-affe-2560992fe39d","added_by":"auto","created_at":"2025-11-04 10:27:08","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":2535999,"visible":true,"origin":"","legend":"supplementary material","description":"","filename":"supplementarymaterial2.0.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7819870/v1/733656794da08515e8adc356.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"DeepPMD: A Comprehensive Deep Learning Framework for Primary-Metastatic Classification and Origin Prediction in Lung Adenocarcinoma – Multi-Center Whole Slide Image Validation","fulltext":[{"header":"Introduction","content":"\u003cp\u003eLungs are among the most common sites for malignant tumor metastasis, with an estimated 20% to 54% of patients with extrathoracic malignancies developing pulmonary metastases during their disease course, resulting in a dramatic decline in prognosis and survival rates.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e For instance, the 5-year survival rate for patients with localized colorectal cancer is 91%, a figure that plummets to 14% following distant metastasis.\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e Therefore, for any pulmonary lesion, accurately identifying the origin of a tumor is a critical prerequisite for guiding subsequent therapy and improving patient outcomes.\u003csup\u003e\u003cspan additionalcitationids=\"CR4 CR5\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e Immunohistochemistry (IHC), valued for its efficiency, accuracy, and affordability, is the primary method for identifying tumor origins. However, its clinical application is complicated by the need not only to exclude second primary lung cancers, which can be morphologically indistinguishable from metastases,\u003csup\u003e7\u003c/sup\u003e but also to address the diagnostic challenge posed by Cancer of Unknown Primary (CUP) .\u003csup\u003e8\u0026ndash;12\u003c/sup\u003e Pathologists, particularly those with limited experience in primary care settings, tend to expand IHC panels to avoid misdiagnosis, which increases medical costs, sample consumption, and patient burden.\u003c/p\u003e\u003cp\u003eThe digitization of pathology through the Whole Slide Images (WSIs) has recently created unprecedented opportunities to apply artificial intelligence (AI). Deep learning models have demonstrated expert-level performance in a range of histopathological tasks, including Gleason grading of prostate cancer biopsies\u003csup\u003e\u003cspan additionalcitationids=\"CR14\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e and predicting survival in colorectal cancer.\u003csup\u003e\u003cspan additionalcitationids=\"CR17\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e This success has extended robustly into pulmonary pathology, where models can now predict genetic mutations,\u003csup\u003e19,20\u003c/sup\u003e patient survival,\u003csup\u003e21\u003c/sup\u003e immunotherapy response,\u003csup\u003e22\u003c/sup\u003e and even future metastasis risk directly from H\u0026amp;E slides.\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e Beyond these applications lies the fundamental diagnostic challenge of identifying a tumor's origin, particularly for Cancer of Unknown Primary (CUP). The landmark TOAD model,\u003csup\u003e24\u003c/sup\u003e firstly established the feasibility of using AI to address this challenge by predicting a primary site from histopathology slides.\u003c/p\u003e\u003cp\u003eHowever, critical limitations remain in the current AI approaches. First, a lack of validation on multi-center, multi-device data casts doubt on their generalizability, as algorithmic performance may not translate across heterogeneous data from different institutions. Second, most models rely solely on morphological information, ignoring crucial clinical data, further their decision-making processes often lack the transparency needed for clinical trust. To address these limitations, we aimed to develop and validate DeepPMD (Deep Learning for Primary-Metastatic Differentiation), a deep learning framework that integrates multi-center WSI data with clinical information to accurately differentiate primary from metastatic tumors in the lung and identify the origin of tumors. We further sought to ensure model transparency through interpretability analysis and rigorously validate its performance in a large, multi-center cohort to assess its real-world clinical utility.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy design and participants\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this national, multi-center and retrospective study, we established DeepPMD to effectively identify primary and metastatic tumors and predict the primary tumor site. To develop and internally validate DeepPMD, we established a discovery cohort of 773 patients (providing 1314 WSIs) from Southern Medical University (SMU) between January 2009 and March 2023 (Figure S1). Then we collected three independent cohorts from three cities throughout China for external validation: External #1 was from the First Affiliated Hospital of Shandong First Medical University, comprising 453 patients (453 WSIs) between 2004 and 2014; External #2 was from The Fourth Hospital of Hebei Medical University, with 547 patients (1007 WSIs) between 2019 and 2024; and External #3 was from The Sixth Affiliated Hospital of Guangzhou Medical University, with 187 patients (318 WSIs) between 2010 and 2023. In total, the external validation set consisted of 1187 patients and 1778 WSIs (Table S1).\u003c/p\u003e\n\u003cp\u003e\u003cspan id=\"_Toc841729566\"\u003e\u0026nbsp;All WSIs were scanned at 20\u0026times; and stored in SVS, TIFF, and MRXS formats and finally converted to SVS format for training. To ensure robust model training and evaluation, pathological types that were present across all four cohorts but had fewer than 2 cases in any single cohort were consolidated into an \u0026quot;Others\u0026quot; category. This resulted in a final classification schema of six categories: lung, gastrointestinal, breast, kidney, liver, and others. Inclusion and exclusion criteria and further details are provided in the appendix (Supplementary Criteria, Supplementary Figure S1).\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee and Institutional Review Board of Nanfang Hospital, Southern Medical University (Approval No. NFEC-2024-379). All experiments were conducted in accordance with the Declaration of Helsinki. Since the study did not involve the collection of private information, we requested a waiver of informed consent for the internal, retrospectively collected data, which was granted by all four participating centres.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Development\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe model was developed and internally validated on 773 patients from the Southern Medical University (SMU) cohort, which were divided at the patient level into a training set (n=541), a test set (n=155), and an internal validation set (n=77) at a 7:2:1 ratio. We selected the H-optimus-1 model, which ranked first in overall performance on lung cancer-related diagnostic tasks and excelled in primary site prediction according to a large-scale benchmark study for pathology foundation models (PathBench) as our foundation model.\u003csup\u003e25\u003c/sup\u003e H-optimus-1 is a vision foundation model with a ViT-G/14 architecture, pre-trained on a dataset of approximately one million WSIs, covering more than 50 tissue types and three scanner devices. To enhance classification performance, we followed the approach of the TOAD model by integrating these patch-level features with patient-level clinicopathological data (gender and age) to create multimodal inputs.\u003c/p\u003e\n\u003cp\u003eOur architecture incorporates two parallel processing branches with gated attention mechanisms: one for primary versus metastatic cancer classification and another for the six-class primary site prediction. This multi-task attention pooling dynamically weights patch features according to their diagnostic relevance for each specific task. These weighted features are aggregated into a comprehensive slide-level representation and combined with the clinical information for the final prediction. Further details on WSI preprocessing and model architecture are available in the Supplementary Methods.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Validation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDeepPMD was evaluated through a multi-stage validation process to assess its accuracy, generalizability, and robustness. The model was firstly validated on the internal validation set, followed by testing on three independent external cohorts to assess real-world generalizability. Furthermore, its robustness was assessed on different specimen types (excisional and aspiration biopsies) within the whole external datasets, and a per-class performance analysis was conducted for each primary origin. Then we validated the clinical relevance and decision-making process of the model via a multi-step interpretability analysis. Firstly, t-SNE was used to visualize the feature space of the learned features from both DeepPMD and TOAD to assess class separability. Next, to evaluate DeepPMD\u0026apos;s attention mechanism, a senior pathologist scored attention heatmaps from 100 randomly selected cases from each of the three external validation cohorts. The scoring was based on the degree of overlap between high-attention regions and actual tumor areas, using a 5-point scale: 1 (0\u0026ndash;20%), 2 (20\u0026ndash;40%), 3 (40\u0026ndash;60%), 4 (60\u0026ndash;80%), and 5 (80\u0026ndash;100%). Finally, to understand the morphological basis of the model\u0026apos;s predictions, high-attention patches from the heatmaps were clinically interpreted by the pathologist based on the pathological criteria for differentiating primary lung adenocarcinoma from metastatic carcinomas. This heatmap scoring and patch-level interpretation were performed only for the DeepPMD model. The performance of DeepPMD was benchmarked against the state-of-the-art TOAD model across all evaluations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModel performance was evaluated for both the binary primary-metastatic classification and the multi-class origin prediction tasks. The overall discriminative ability of the models was assessed using the area under the receiver operating characteristic curve (AUC). We also reported accuracy (ACC) and the F1 score, which considers both precision and recall, to provide a comprehensive evaluation. To account for class imbalance in the multi-class task, we reported macro-averaged (the unweighted mean of per-class scores) and weighted-averaged (weighted by per-class support) F1 scores. The stability of all performance metrics was quantified by 95% confidence intervals (CIs), generated using non-parametric bootstrapping with 1,000 resamplings. To compare the performance between the DeepPMD and TOAD models, the two-sided Wilcoxon signed-rank test was used, with a p-value of less than 0.05 considered statistically significant.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eIn the SMU cohort, 1347 WSIs were available from 773 individuals (427 [53.85%] men and 366 [46.015%] women) (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB). Median age was 58 years [IQR 18\u0026ndash;86]. For the external validation cohorts, we enrolled 1778 WSIs from 1187 patients across three medical centers around from north to south in China (701 [59.06%] men and 486 [40.94%] women; median age 61 years [IQR 19\u0026ndash;86]). The cohorts exhibited significant differences in biopsy types and technical parameters, with aspiration biopsies being predominant in the internal cohort (45.67%), while excisional biopsies were more common in the external cohort (59.06%). Additionally, different brands of WSI scanners (Leica, Ibingli, and 3DHISTECH) were used, ensuring the diversity of the dataset for a more comprehensive evaluation of the model's generalization capability.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFor the fundamental performance of the model, DeepPMD accurately differentiated a primary tumor from a metastatic tumor and identifying the origin of tumor (Macro-AUC 0.993 [95%CI 0.984, 0.999]), and identified the origin of each metastatic tumor (Macro-AUC 0.992 [95%CI 0.985, 0.998]) in the whole internal validation set (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). On each external cohort, the value of Macro-AUC was from 0.850 (95%CI 0.813, 0.884) to 0.981 (95%CI 0.962, 0.995) for differentiation between primary and metastatic cancer, and from 0.914 (95% CI 0.889, 0.935) to 0.981(95% CI 0.823, 0.996) for primary tumor site prediction (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAnd on the pooled external validation cohort, DeepPMD achieved stable and excellent performance regardless of specimen types (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA-B). On excisional biopsy samples, which typically provide more comprehensive tissue architecture, DeepPMD demonstrated exceptional performance. It had a Macro-AUC value of 0.962 (95%CI 0.950, 0.973), Macro-ACC value of 0.889 (95%CI 0.869, 0.910) and Macro-F1 value of 0.837 (95%CI 0.808, 0.865) for discrimination of primary versus metastatic, while a Macro-AUC value of 0.974 (95%CI 0.949, 0.991), Macro-ACC value of 0.850 (95%CI 0.747, 0.925) and Macro-F1 value of 0.767 (95% CI 0.659, 0.827) for primary tumor site prediction. In the more challenging aspiration biopsies, the model had a Macro-AUC value of 0.944 (95%CI 0.920, 0.966), Macro-ACC value of 0.844 (95%CI 0.801, 0.883) and Macro-F1 value of 0.841 (95%CI 0.799, 0.877) for discrimination of primary versus metastatic, while a Macro-AUC value of 0.935 (95%CI 0.913, 0.953), Macro-ACC value of 0.728 (95%CI 0.670, 0.781) and Macro-F1 value of 0.756 (95% CI 0.702, 0.801) for primary tumor site prediction.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eMeanwhile, recognizing the imperative for high reliability in clinical applications and the potential for macro-averaging to obscure poor performance in minority classes, we conducted a stratified analysis of the model's primary site prediction performance on aspiration versus excisional biopsy samples from all the external cohorts (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC). On aspiration biopsy samples, the DeepPMD model demonstrated excellent performance across all cancer types, with particularly strong results in predicting gastric \u0026amp; colorectal cancer (AUC\u0026thinsp;=\u0026thinsp;0.965, 95% CI 0.947\u0026ndash;0.979) and renal cancer (AUC\u0026thinsp;=\u0026thinsp;0.966, 95% CI 0.924\u0026ndash;0.992). The model similarly exhibited a high level of accuracy on excisional biopsy samples, achieving remarkable discrimination for renal cancer with an AUC of 0.999 (95% CI 0.996\u0026ndash;1.000). In a combined analysis of both biopsy types, the DeepPMD model maintained its robust predictive capabilities across all primary sites, further demonstrating its generalizability and reliability, irrespective of the biopsy acquisition method.\u003c/p\u003e\u003cp\u003eWhen benchmarked against the TOAD model, DeepPMD demonstrated significantly higher performance (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001 for all key comparisons). On the internal validation set, DeepPMD's Macro-AUC advantage over TOAD was modest (10.8% for the binary task and 12.1% for the six-class task). However, this gap widened substantially on external cohorts, with the margin for the binary task on External #1 increasing to 31.2% (0.850 vs 0.648). TOAD's poor generalization was most pronounced in the six-class prediction task, where its Macro-ACC only dropped to 0.269 on External #1, and on aspiration biopsies, where its Macro-AUC (0.731) was markedly lower than DeepPMD's (0.935) .\u003c/p\u003e\u003cp\u003eThis performance advantage stems from DeepPMD\u0026rsquo;s ability to learn more discriminative features, as visualized by t-SNE dimensionality reduction (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, Supplementary Figure S2). The DeepPMD model formed well-demarcated feature clusters for both binary (primary vs. metastatic) and six-class (primary site prediction) tasks on the combined dataset. This clear clustering was consistently reproduced in the internal and three independent external validation sets, confirming the robust cross-center generalizability of the learned features.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn the t-SNE visualization of the combined dataset, tumors from different origins formed dense and distinct clusters. Notably, the clusters for lung adenocarcinoma (LUAD) and breast cancer exhibited some overlap, which is primarily attributable to their shared histopathological similarities, such as glandular architectures, abundant eosinophilic cytoplasm, and comparable nuclear atypia. This observation was corroborated by the model's confusion matrix: across all samples, 56 LUAD cases were misclassified as breast cancer, and 21 breast cancer cases were misclassified as LUAD (Supplementary Figure S3). This demonstrates that DeepPMD's confusion is not random error but reflects a learned morphological discrimination logic analogous to that of human pathologists.\u003c/p\u003e\u003cp\u003eIn contrast, the feature space of the TOAD model appeared disorganized. Particularly in the external validation sets, feature points from different classes were severely intermixed, lacking discernible cluster structures and clear boundaries. This evidence strongly indicates that DeepPMD learns efficient and robust discriminative features that successfully generalize to heterogeneous, multi-center data, whereas the TOAD model fails to learn effective universal features.\u003c/p\u003e\u003cp\u003eTo evaluate the model's interpretability, attention heatmaps from all external validation cohorts were systematically scored by a senior pathologist (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). The results showed that the proportions of cases with high-focus scores (4 or 5) were 86.28% in External #1, 90.30% in External #2, and 89.84% in External #3. This demonstrates the model's consistent and accurate localization ability across datasets from different medical centers. Then we analyzed its high-attention regions (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). Analysis of high-attention patches confirmed the model's decisions were based on key histomorphological criteria: for primary lung adenocarcinoma, the model focused on areas with mixed growth patterns, particularly the highly specific lepidic growth pattern; for metastatic cancers, its attention was origin-specific, identifying the classic triad of serrated glands, nuclear palisading, and \u0026ldquo;dirty\u0026rdquo; necrosis in colorectal metastases; locating the single-file line (Indian file) arrangements characteristic of lobular breast carcinoma; concentrating on thick trabecular structures of eosinophilic polygonal cells separated by a rich sinusoidal network in hepatocellular carcinoma; and capturing nests of clear cells within a dense, thin-walled capillary network in clear cell renal carcinoma. This visual evidence strongly suggests that the model has learned to recognize the tissue-specific morphological cues that are critical for differential diagnosis, rather than relying on artifacts or spurious features, thereby confirming the pathological basis of its decisions.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFinally, to assess the model's potential utility in addressing the clinical challenge of Cancer of Unknown Primary (CUP), we designed a consistency analysis(Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA). In this analysis, the model's top-k predictions were compared against the k differential diagnoses provided clinically; an overlap of over 50% was considered consistent. The analysis revealed that DeepPMD's predictions achieved a consistency score of 0.86 with the clinical diagnostic path, whereas the TOAD model scored only 0.37. Based on this performance, we developed SmartPath, a system that generates a probability-ranked list of potential primary sites for pathologists directly from WSIs (Supplementary Figure S4). This approach optimizes the diagnostic workflow by guiding the selection of targeted IHC panels, thereby shortening diagnostic turnaround time, increasing precision, and conserving tissue, particularly in cases with limited clinical data or tissue samples\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed and validated DeepPMD, a deep learning model that accurately differentiates between primary and metastatic lung tumors and predicts their primary origin using only standard WSIs and basic clinical data. Evaluated on a large-scale cohort from four independent medical centers using three different scanners, DeepPMD demonstrated exceptional accuracy and robust generalizability, significantly outperforming existing benchmark models. Furthermore, our interpretability analysis revealed that the model's diagnostic logic is highly concordant with the histopathological criteria used by pathologists, thereby reinforcing its reliability for clinical decision-making.\u003c/p\u003e\u003cp\u003eIn clinical practice, when faced with a pulmonary tumor of unknown origin, pathologists rely on patient history and histomorphological features to select a broad IHC panel to ascertain its primary or metastatic nature and potential origin. In complex cases or with less experienced pathologists, a more extensive IHC panel is often chosen, which prolongs diagnostic turnaround time, consumes limited biopsy tissue, and increases healthcare costs. DeepPMD is poised to optimize this diagnostic workflow. By inputting a WSI into the system interface, pathologists can instantaneously receive a probability-ranked list of potential primary sites, enabling the selection of a more targeted, narrow-spectrum IHC panel. This approach can significantly shorten the diagnostic timeline, enhance precision, and conserve valuable time for therapeutic planning, proving especially beneficial in cases with limited clinical history or scant tissue samples. Moreover, as conventional broad-panel IHC can sometimes yield inconclusive results, leading to treatment delays and potential disputes, our model provides an objective, data-driven reference that reduces diagnostic uncertainty, supports the final pathological conclusion, and may mitigate risks associated with diagnostic ambiguity.\u003c/p\u003e\u003cp\u003eThe robustness of the model is fundamental to its clinical applicability. By validating DeepPMD across four centers with different WSI scanners and clinical practices, we have shown that the morphological features it learns are robust and generalizable. Furthermore, our interpretability analyses confirmed that the model's diagnostic logic aligns closely with the visual criteria used by pathologists, enhancing its trustworthiness for clinical decision-making.\u003c/p\u003e\u003cp\u003eDespite these encouraging results, our study has several limitations. First, the current model was trained and validated on a selection of the most common metastatic tumor types, and its performance on rarer cancer origins has yet to be explored. Second, this was a retrospective study. Although rigorously validated on multi-center data, a prospective study has not yet been conducted to evaluate its real-world impact on clinical workflows, such as direct effects on treatment decisions. Future work will focus on expanding the range of tumor types and validating the clinical utility of DeepPMD in a prospective trial.\u003c/p\u003e\u003cp\u003eIn conclusion, DeepPMD demonstrates immense potential as an accurate, efficient, and interpretable AI-powered tool for the challenging task of tumor origin identification. By optimizing IHC testing strategies, it promises to shift the paradigm in pathological diagnosis, ultimately benefiting patients.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eContributors\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eS.G.: Conceptualization, Data Curation, Validation, Methodology, Visualization, Writing \u0026ndash; Original Draft, Writing \u0026ndash; Review \u0026amp; Editing. J.M.: Methodology, Software, Formal Analysis,Validation, Visualization. J.L.: Conceptualization, Methodology, Data Curation, Validation, Investigation. C.Z.: Validation, Investigation. D.L., X.W.: Writing \u0026ndash; Review \u0026amp; Editing. F.Z., Y.X., Z.G.: Visualization. Z.Z., Y.T., A.A., W.Q., X.Z., B.L., Q.X., Z.W.: Data Curation. Y.W., Y.L., J.C., H.C., L.L.: Supervision, Project Administration, Funding Acquisition. L.L., Y.W., Y.L., J.C. also provided: Resources. H.C. also contributed to: Software. All authors reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDatasets are not publicly available due to patient privacy considerations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe underlying code for this study is publicly available in the GitHub repository (https://github.com/birkhoffkiki/DeepPMD).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by grants from the National Natural Science Foundation of China (82273358), Chongqing Technology innovation and application development special major project (CSTB2024TIAD-STX0003).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eBudczies J, von Winterfeld M, Klauschen F, et al. The landscape of metastatic progression patterns across major human cancers. Oncotarget. 2015;6(1):570-83.\u003c/li\u003e\n \u003cli\u003eSiegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin. 2024;74(1):12-49. doi:10.3322/caac.21820\u003c/li\u003e\n \u003cli\u003eChou TY, Dacic S, Wistuba I, et al. Differentiating separate primary lung adenocarcinomas from intrapulmonary metastases with emphasis on pathological and molecular considerations: recommendations from the International Association for the Study of Lung Cancer Pathology Committee. J Thorac Oncol. 2025;20(3):311-30. doi:10.1016/j.jtho.2024.11.016\u003c/li\u003e\n \u003cli\u003eChen K, Liu A, Wang C, et al. Multidisciplinary expert consensus on diagnosis and treatment of multiple lung cancers. Med. 2025;6(4):100643. doi:10.1016/j.medj.2025.100643\u003c/li\u003e\n \u003cli\u003eBrims F, McWilliams A, Williamson J, Siemienowicz M, Leong TL; Thoracic Society of Australia \u0026amp; New Zealand Lung Cancer Working Party. The TSANZ Practical Guide for Clinicians in the Management of Screen- and Incidentally-Detected Nodules. Respirology. 2025;30(7):558-73. doi:10.1111/resp.70065\u003c/li\u003e\n \u003cli\u003eAnagnostopoulos AK, Gaitanis A, Gkiozos I, et al. Radiomics/radiogenomics in lung cancer: basic principles and initial clinical results. Cancers (Basel). 2022;14(7):1657. doi:10.3390/cancers14071657\u003c/li\u003e\n \u003cli\u003eChuang SC, Sc\u0026eacute;lo G, Lee YC, et al. Risks of second primary cancer among patients with major histological types of lung cancers in both men and women. Br J Cancer. 2010;102(7):1190-5. doi:10.1038/sj.bjc.6605616\u003c/li\u003e\n \u003cli\u003eKato S, Alsafar A, Walavalkar V, Hainsworth J, Kurzrock R. Cancer of unknown primary in the molecular era. Trends Cancer. 2021;7(5):465-77. doi:10.1016/j.trecan.2020.11.002\u003c/li\u003e\n \u003cli\u003eKr\u0026auml;mer A, Bochtler T, Pauli C, et al. Cancer of unknown primary: ESMO Clinical Practice Guideline for diagnosis, treatment and follow-up. Ann Oncol. 2023;34(3):228-46. doi:10.1016/j.annonc.2022.11.013\u003c/li\u003e\n \u003cli\u003eEconomopoulou P, Mountzios G, Pavlidis N, Pentheroudakis G. Cancer of unknown primary origin in the genomic era: elucidating the dark box of cancer. Cancer Treat Rev. 2015;41(7):598-604. doi:10.1016/j.ctrv.2015.05.010\u003c/li\u003e\n \u003cli\u003eBoscolo Bielo L, Belli C, Crimini E, et al. Cancers of unknown primary origin: real-world clinical outcomes and genomic analysis at the European Institute of Oncology. Oncologist. 2024;29(6):504-10. doi:10.1093/oncolo/oyae038\u003c/li\u003e\n \u003cli\u003eSpurgeon L, Mitchell C, Cook N, Conway AM. Cancer of unknown primary: the hunt for its elusive tissue-of-origin - is it time to call off the search?. Br J Cancer. Published online July 4, 2025. doi:10.1038/s41416-025-03073-7\u003c/li\u003e\n \u003cli\u003eNagpal K, Foote D, Tan F, et al. Development and validation of a deep learning algorithm for Gleason grading of prostate cancer from biopsy specimens. JAMA Oncol. 2020;6(9):1372-80. doi:10.1001/jamaoncol.2020.2485\u003c/li\u003e\n \u003cli\u003eBulten W, Kartasalo K, Chen PC, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nat Med. 2022;28(1):154-63. doi:10.1038/s41591-021-01620-2\u003c/li\u003e\n \u003cli\u003eCampanella G, Hanna MG, Geneslaw L, et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med. 2019;25(8):1301-9. doi:10.1038/s41591-019-0508-1\u003c/li\u003e\n \u003cli\u003eMahajan A, Kania V, Agarwal U, et al. Deep-learning-based predictive imaging biomarker model for EGFR mutation status in non-small cell lung cancer from CT imaging. Cancers (Basel). 2024;16(6):1130. doi:10.3390/cancers16061130\u003c/li\u003e\n \u003cli\u003eSkrede OJ, De Raedt S, Kleppe A, et al. Deep learning for prediction of colorectal cancer outcome: a discovery and validation study. Lancet. 2020;395(10221):350-60. doi:10.1016/S0140-6736(19)32998-8\u003c/li\u003e\n \u003cli\u003eFoersch S, Glasner C, Woerl AC, et al. Multistain deep learning for prediction of prognosis and therapy response in colorectal cancer. Nat Med. 2023;29(2):430-9. doi:10.1038/s41591-022-02134-1\u003c/li\u003e\n \u003cli\u003eCoudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med. 2018;24(10):1559-67. doi:10.1038/s41591-018-0177-5\u003c/li\u003e\n \u003cli\u003eBabu B, Singh J, Salazar Gonz\u0026aacute;lez JF, et al. A narrative review on the role of artificial intelligence (AI) in colorectal cancer management. Cureus. 2025;17(2):e79570. doi:10.7759/cureus.79570\u003c/li\u003e\n \u003cli\u003ePan X, Lin H, Han C, et al. Computerized tumor-infiltrating lymphocytes density score predicts survival of patients with resectable lung adenocarcinoma. iScience. 2022;25(12):105605. doi:10.1016/j.isci.2022.105605\u003c/li\u003e\n \u003cli\u003eRakaee M, Tafavvoghi M, Ricciuti B, et al. Deep learning model for predicting immunotherapy response in advanced non-small cell lung cancer. JAMA Oncol. 2025;11(2):109-18. doi:10.1001/jamaoncol.2024.5356\u003c/li\u003e\n \u003cli\u003eZhou H, Watson M, Bernadt CT, et al. AI-guided histopathology predicts brain metastasis in lung cancer patients. J Pathol. 2024;263(1):89-98. doi:10.1002/path.6263\u003c/li\u003e\n \u003cli\u003eLu MY, Chen TY, Williamson DFK, et al. AI-based pathology predicts origins for cancers of unknown primary. Nature. 2021;594(7861):106-10. doi:10.1038/s41586-021-03512-4\u003c/li\u003e\n \u003cli\u003eMa J, Xu Y, Zhou F, et al. PathBench: a comprehensive comparison benchmark for pathology foundation models towards precision oncology. arXiv. 2025 May 26. doi:10.48550/arXiv.2505.20202\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Tumor Origin Prediction, Computational Pathology, Cancer of Unknown Primary (CUP), Deep learning, Whole Slide Images","lastPublishedDoi":"10.21203/rs.3.rs-7819870/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7819870/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eAmong solid malignancies, metastatic spread accounts for ~\u0026thinsp;90% of cancer-related deaths. For unknown pulmonary lesions, accurate diagnosis is crucial for treatment. In diagnostically challenging cases like second primary tumor and cancer of unknown primary (CUP), traditional morphological analysis and immunohistochemistry (IHC) often remain inconclusive, delaying critical therapeutic intervention.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe developed a deep learning model (DeepPMD) that makes predictions using H\u0026amp;E stained whole slide images and basic clinical data. The model was trained and validated on a single-center cohort of 793 patients, then externally validated on three additional cohorts totaling 1187 patients from independent centers.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eDeepPMD achieved accurate prediction of tumor nature and site of origin, outperforming benchmarks in all tests. In external validation, the model achieved a macro-AUC of 0.974 (95% CI 0.949, 0.991) for origin prediction on excisional biopsies and 0.935 (95% CI 0.913, 0.953) on aspiration biopsies. Its diagnostic logic aligned with pathological criteria, and its CUP predictions showed high concordance with clinical predictions (consistency score 0.86).\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e\u003cp\u003eDeepPMD is an accurate, generalizable, and interpretable AI tool for identifying pulmonary tumor origins from WSIs. It can provide probability-ranked potential primary sites to guide targeted IHC testing, optimizing diagnostic workflow, shortening turnaround times, and conserving tissue samples.\u003c/p\u003e","manuscriptTitle":"DeepPMD: A Comprehensive Deep Learning Framework for Primary-Metastatic Classification and Origin Prediction in Lung Adenocarcinoma – Multi-Center Whole Slide Image Validation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-04 10:27:03","doi":"10.21203/rs.3.rs-7819870/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-communications","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"NCOMMS","sideBox":"Learn more about [Nature Communications](http://www.nature.com/ncomms/)","snPcode":"","submissionUrl":"https://mts-ncomms.nature.com/","title":"Nature Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Communications","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"cc6e2579-821e-4b4b-b4fc-95f182af6c28","owner":[],"postedDate":"November 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":56571842,"name":"Biological sciences/Cancer/Lung cancer/Small-cell lung cancer"},{"id":56571843,"name":"Biological sciences/Computational biology and bioinformatics/Software"}],"tags":[],"updatedAt":"2025-11-04T10:27:03+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-04 10:27:03","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7819870","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7819870","identity":"rs-7819870","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}