Predicting metastatic potential of primary cutaneous melanomas utilizing weakly supervised vision language model

Predicting metastatic potential of primary cutaneous melanomas utilizing weakly supervised vision language model | 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 Predicting metastatic potential of primary cutaneous melanomas utilizing weakly supervised vision language model Filip Dahlén, Ivan Shujski, Filmon Yacob, Ida Häggström, Jovana Jovanovic, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6297243/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 01 Apr, 2026 Read the published version in Scientific Reports → Version 1 posted 12 You are reading this latest preprint version Abstract Cutaneous melanoma is an aggressive form of skin cancer. Knowledge if a primary melanoma is likely to metastasize is crucial for treatment and survival prediction of melanoma patients. We aimed to develop a predictive tool for determining metastatic potential in primary melanomas utilizing a weakly supervised vision language model. A total of 426 routine stained whole slide images (WSI), along with corresponding histopathological features (Breslow thickness, diameter, presence of dermal mitoses, ulceration and regression), were collected. Of these, 341 samples were used for training and validation, while 85 were reserved as a holdout test set. WSIs were split into patches, and feature embeddings were extracted using Prov-GigaPath. Histopathological features were converted to text, with embeddings generated by BiomedBERT. We developed a multimodal transformer integrating WSIs and histopathological features and conducted an ablation study comparing it to (1) TransMIL using only WSIs and (2) an MLP using only histopathological features. Each model employed a bagging ensemble with five cross-validation models. The multimodal transformer achieved an AUC of 0.887, slightly higher than TransMIL (0.883) and notably better than BertMLP (0.800), highlighting the benefit of including imaging and clinical data for early recognition of melanomas with high metastatic potential. Health sciences/Diseases/Cancer Health sciences/Diseases/Cancer/Skin cancer/Melanoma Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Health care Health sciences/Medical research Physical sciences/Mathematics and computing/Computer science WSI Transformer BioMedBERT Melanoma Prov-Gigapath Foundation models Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Cutaneous melanoma is one of the deadliest skin cancers and a growing cause of death and morbidity worldwide. The prognosis for stage III and IV melanoma patients is significantly worse compared to patients in stage I or II without metastatic disease 1 . Thus, adjuvant treatments with both targeted therapies and immunotherapy are often considered for patients in stage III and IV 2 . There is evidence indicating that metastases from a primary cutaneous melanoma often occur long before the melanoma is diagnosed 2 . However, some melanomas metastases are detected early, while some melanomas stay in a “dormant” stage for years 3 . Melanoma metastasizes lymphogenously (to regional lymph nodes or as in-transit metastases), but also hematogenously with distant metastases 1 , 4 . Thus, knowing if a melanoma is likely to metastasize is crucial for treatment and survival of melanoma patients. Primary tumors contain morphological markers that may indicate metastatic potential. At the time of a primary melanoma diagnosis, it is possible to use clinico-pathological parameters to predict the likelihood of sentinel lymph node metastases 5 and of eventual death from melanoma 6 , 7 . The most informative prognostic features for survival are the Breslow thickness, ulceration and presence of dermal mitoses 6 . Current diagnostic practices primarily involve histopathological evaluation, which can be labor-intensive and subject to interobserver variability. Despite the progress of self-supervised learning and foundation models in computer vision and natural language processing, their application in the medical domain remains in its early stages. One of the main reasons is the limited availability of publicly accessible data compared to other domains — a challenge particularly pronounced in pathology due to the relatively slow adoption of digital pathology and the inevitable time constraints in collecting patient data. However, recent efforts have led to the public release of several foundation models pre-trained on large-scale clinical datasets. These developments have significantly advanced computational pathology research, by lowering the barriers for smaller research groups and accelerating the translation of AI-based methods into clinical practice. Campanella et al. 8 made a clinical benchmark of public self-supervised pathology foundation models and measured the performance of the models on two types of downstream tasks, disease detection and computational biomarkers. In general, the models trained using DINO 9 and DINOv2 10 (SP21M, SP85M, UNI, Virchow, and Prov-GigaPath) achieved comparable performance and in biomarker prediction UNI and Prov-Gigapath were just as good or better than the other models with Prov-Gigapth performing slightly better than UNI. In biomarker prediction, an overall trend towards higher performance with larger models was also observed. 8 These emerging self-supervised models provide a promising starting point for further advancements in computational pathology. Alongside these developments, recent developments in computational pathology have enabled development of prognostic models based on digitized routine stained histopathological whole slide images (WSIs) 11 . Significant progress has been made in detecting and classifying cancers and identifying metastases in lymph nodes using machine learning models. In the domain of histological imaging, both Vision Transformers 12 and more traditional convolutional neural network (CNN) architectures have demonstrated significant success in a range of applications, such as detecting breast cancer metastases and classifying cancer subtypes in lung, kidney and colorectal tissues 13 – 18 . Furthermore, transformer based models, CNN models and MIL models have been particularly effective in detection of metastases 16 , 19 – 23 . Despite these advancements, only a limited number of studies have focused on identifying metastatic potential in primary tumors including Brinker et al. 24 which predicted sentinel lymph node metastatic status directly from routine histology of primary melanoma tumors using a CNN-based approach. Similarly, Knuutila et al. 25 investigated metastatic primary cutaneous squamous cell carcinoma instead using a residual neural network (RNN) architecture. In another study, Kulkarni et al. 26 analyzed primary melanoma tumors to identify patients at risk for visceral recurrence and death using a CNN combined with an RNN architecture. In this study, we aimed to predict metastatic risk in primary melanomas by developing a multimodal transformer trained on both image and text embeddings referred to as MultiTrans. Furthermore, we aimed to conduct an ablation study to compare this model to TransMIL 27 trained solely with image embeddings and a MLP trained only with text embeddings referred to as BertMLP. Results MultiTrans achieved an AUC of 0.887, an accuracy of 0.871 with a sensitivity of 0.921 and a specificity of 0.794 when evaluated at the Youden index. The mean receiver operating characteristic (ROC) curve is visualized in Fig 1, the confusion matrix at the Youden index is visualized in Fig 2. To further interpret the model’s predictions, heatmaps were used to highlight regions in whole slide images (WSI) that contributed most to classification. Figure 3 displays tumor regions in three different samples correctly identified by MultiTrans. Ablation Study To evaluate the impact of different model architectures, we compared MultiTrans, TransMIL, and BertMLP on the hold-out test set. TransMIL The TransMIL as described by Shao et al. 27 was used. The TransMil model, given a bag of image embeddings, will perform multi-head self-attention 38 on the input embeddings to capture the correlation between the different embeddings. Furthermore, the attended features are fed to a Pyramid Position Encoding Generator (PPEG) that will encode the spatial information before being fed to the last Transformer layer to aggregate the morphological information. Finally, the aggregated features are fed to a single MLP layer for classification. BertMLP Classification of text embeddings was performed using a MLP consisting of three layers with ReLU activation functions in between. The full performance metrics are summarized in Table 1. The mean ROC curve for each model is visualized in Fig S1, while confusion matrices evaluated at the Youden index are shown in Fig S2. MultiTrans achieved the highest accuracy (0.871) compared to both TransMIL (0.856) and BertMLP (0.729) when evaluated at the Youden index. Both MultiTrans and TransMIL demonstrated high AUC and sensitivity, indicating strong ability to identify metastatic cases with MultiTrans performing slightly better (0.921 vs 0.882), detecting more metastatic cases than TransMIL (47 TP vs 45 TP). However, TransMIL achieved higher specificity (0.824 vs 0.794) indicating it reduced false positives more effectively (6 FP vs 7 FP). The AUC for MultiTrans and TransMIL were similar with slightly higher values for MultiTrans (0.887 vs 0.883). MultiTrans and TransMIL significantly outperformed BertMLP across all metrics besides Specificity where BertMLP had higher value than MultiTrans (0.882 vs 0.794), meaning it produced fewer false positives than MultiTrans (4 FP vs 7 FP). MultiTrans minimized false negatives (FN = 4), which is crucial for ensuring metastatic cases are not missed. BertMLP had the highest false negatives (FN = 19), making it unreliable for detecting metastatic cases. These results suggest that WSIs alone provided sufficient information for accurate predictions, while the inclusion of histopathological features further enhances model performance. Table 1 : Performance metrics for MultiTrans, TransMIL and BertMLP. Model Accuracy Sensitivity Specificity AUC TP FP FN TN MultiTrans 0.871 0.921 0.794 0.887 47 7 4 27 TransMIL 0.856 0.882 0.824 0.883 45 6 6 28 BertMLP 0.729 0.627 0.882 0.802 32 4 19 30 Discussion In this paper, we predicted metastatic potential in primary melanomas by comparing a TransMIL trained on histopathological image data, MultiTrans trained on image data together with clinical data, and a BertMLP trained on clinical data. The results show that TransMIL is able to detect early signs of metastatic potential in the primary tumors with relatively high accuracy. MultiTrans achieves a slightly higher AUC compared to TransMIL but not large enough to be significant. Both TransMIL and MultiTrans outperform BertMLP. The results highlight the impact of using image data for this prediction task. Brinker et al. 24 predicted the sentinel lymph node metastatic status directly from routine histology of primary melanoma tumors, achieving an AUC of 0.618 when using WSIs alone, 0.616 when using only histopathological features, and a slightly lower AUC of 0.613 when combining WSIs and histopathological features. Similarly, Knuutila et al. 25 investigated metastatic primary cutaneous squamous cell carcinoma using a residual neural network architecture, reporting an AUC of 0.747 when analyzing WSIs alone, 0.804 when using only histopathological features, and an AUC of 0.917 when developing a risk factor model (RFM) consisting of AI-based WSI predictions, Clark’s level 5, and tumor diameter ≥ 40 mm as risk factors. In another study, Kulkarni et al. 26 analyzed primary melanoma tumors to identify patients at risk of visceral recurrence and death. They employed a CNN combined with a RNN architecture, achieving an AUC of 0.905 and 0.880 in two independent validation sets. In the study by Brinker et al. 24 , tumor regions were first identified and annotated as regions of interest for each WSI, followed by division into patches, which were processed through a CNN and subsequently classified using an MLP layer. Similarly, Knuutila et al. 25 divided each WSI into patches, assigned binary tumor labels based on prior annotations, and further labeled them according to metadata as indicating rapid metastases or not. The patches were then classified using a ResNet-18 architecture. Kulkarni et al. 26 pre-processed the WSIs to isolate tumor regions using QuPath digital pathology software. In contrast to these approaches, which relied on detailed annotations or patch-level labeling, our method is weakly supervised, with no predefined annotations or patch-level labels. Instead, the transformer model learned to identify discriminative regions through self-attention over all patches within each WSI. Additionally, none of the previous studies utilized a foundation model for feature extraction as in our approach, where feature extraction is decoupled from the classification task. The findings from our study suggest that WSIs alone provided sufficient information for accurate predictions, while the inclusion of histopathological features further enhances model performance. These results align with the prior study by Knuutila et al. 25 , which demonstrated improved predictive performance when combining image-based predictions with clinical risk factors. Our study also highlights the importance of utilizing a foundation model to extract relevant features from the image patches together with a transformer to direct attention towards the most relevant regions which is supported by our results being an AUC of 0.883 compared to 0.618 and 0.747 26 . However, the observation that the combined use of WSIs and histopathological features yielded only a marginal improvement over WSIs alone may indicate that the applied data fusion strategy was suboptimal. The approach used in this study may not have effectively aligned or balanced the image embeddings with the histopathological features, potentially limiting the model’s overall performance. Future research should explore more advanced data fusion methods and the development of multimodal embeddings to better integrate different data sources and improve predictive performance. Evidence suggests that melanoma metastases often occurred many months before a primary melanoma diagnosis is made 3 which makes it crucial to identify these aggressive tumors in an early stage. Metastases to regional lymph nodes via lymphatics is the most common form of spread, with around 50% of those who develop metastases having nodal disease as their first site of clinically-detected recurrence 1 , 4 . However, metastases in a distant organ with no evidence of previous or current lymph node disease is seen in about 30% of those who develop metastatic melanoma 29 , 30 suggesting dissemination of malignant cells exclusively via the bloodstream. Lymph node metastases are often diagnosed earlier than metastases at distant sites, with a median interval of 16 months between primary diagnosis and the detection of nodal metastases in one study 29 , while distant metastases tend to be detected a median of 25–40 months after primary diagnosis 31 . In our study of 426 tumors, 249 were metastatic, and of those 128 were in stage III showing lymph node or local in transit metastases while 121 showed distal metastases (stage IV). Despite the three year follow up and sentinel node biopsies conducted in all the patients in the non-metastatic group, some of the non-metastatic group may have developed undetected micrometastatic disease. Furthermore, occasionally melanomas show very late clinical appearance of metastases, sometimes more than 10 years after the primary melanoma was excised 32 . One limitation in our study is the dataset size. A dataset of 426 WSIs, while valuable, may not capture the full variability of tumor morphology and metastatic patterns present in a broader population. This limitation could potentially hinder the generalizability of the model to unseen data or rare cases. The dataset size plays a critical role in improving model performance 33 . Furthermore, it has been shown that simultaneously scaling the dataset size and the model size can lead to significant performance gains. Another limitation is that we didn’t predict disease-specific survival. Additionally, not all WSI had a complete set of histopathological features which most possibly lowered the performance of the models. Also, every WSI had an identification number written on the glass when scanned that could potentially influence the performance, but according to the attention maps, no attention was laid on these areas. We evaluated two sets of text sentences for BertMLP: one including gender and age information and one without, which represents the current approach. The results showed that incorporating gender and age led to lower performance compared to the current approach, which contrasts with the findings of Mervic et al. 31 . One possible explanation for this discrepancy is the imbalance in the dataset, where males outnumber females, potentially introducing bias. However, this warrants further investigation To conclude, TransMIL trained solely on image data was able to detect early signs of metastatic potential in the primary tumors with high accuracy outperforming the BertMLP trained with the histopathological parameters which are the current prognostic standard. MultiTrans combining the text embeddings of histopathological parameters and image data achieved a slightly higher AUC compared to TransMIL and somewhat higher accuracy when evaluated at the Youden index. The results highlight the impact of using image data for this prediction task and the potential in combining image data together with histopathological data and demonstrate high accuracy for early recognition of melanomas with high metastatic potential. Predicting early signs of metastatic potential from primary tumors could enable early targeted treatments for this patient group. Methods An overview of the method is shown in Fig. 4 . Dataset The data was retrospectively collected at the Sahlgrenska University Hospital between 2016–2023. Inclusion criteria were a primary cutaneous malignant melanoma stage I-IV with information about possible metastases and in case of non-metastatic tumor, a negative sentinel node examination and a minimum of 3 years follow-up time. The exclusion criteria were patient cases with metastatic malignant melanoma having more than one primary malignant melanoma tumor to avoid uncertainty of which was the primary tumor that metastasized. One glass slide per tumor harboring the largest Breslow thickness was collected and scanned unidentified using a scanner NanoZoomer S360 Hamamatsu at 40X magnification. The complete dataset consisted of 426 WSIs representing 426 primary melanomas (249 metastatic and 177 non-metastatic) from 425 patients (one patient had two non-metastatic melanomas), detailed in Table 2 . The size of the WSI ranged from 1.4GB up to 5.8GB with dimensions between (157 440 x 55 296) and up to (215 040 x 109824) pixels. The total size of the dataset was around 1.4TB. Table 2 Distribution of clinical and histopathological features across metastatic and non-metastatic classes. Breslow thickness and diameter are presented as mean (min - max) values in millimeters. Mitoses, ulceration, and regression are reported as the fraction of occurrences (present/not present) within each class. All included Total Met Non-met Training and validation set Total Met Non-met Hold-out test set Total Met Non-met Slides WSI 426 249 177 341 198 143 85 51 34 Patches 2 838 864 1 659 336 1 179 528 2 272 424 1 319 472 952 952 566 440 339 864 226 576 Gender Male 231 138 93 182 107 75 49 31 18 Female 195 111 84 159 91 68 36 20 16 Age (Mean, IQR) Male 66 (29–92) 66 (32–92) 65 (29–90) 65 (29–92) 65 (32–92) 65 (29–89) 68 (42–90) 68 (43–89) 67 (42–90) Female 65 (17–94) 67 (21–94) 63 (17–92) 65 (17–94) 67 (21–94) 63 (17–92) 64 (27–94) 64 (27–94) 65 (32–86) Stage III - 128 - - 104 - - 24 - IV - 121 - - 94 - - 27 - histopathological features Breslow (mean, std, p-value) 3.7 (0.3–35) 4.9 (0.5–35) 1.9 (0.3–9.5) 3.5 (0.5–20) 4.6 (0.5–20) 1.9 (0.8–9.5) 4.9 (0.3–35) 6.3 (1.1–35) 1.8 (0.3–0.7) p-value \(\:{1.70\cdot\:10}_{}^{-23}\) Diameter (mean, std) 14.3 (3.0–80.0) 16.3 (3.5–80) 11.6 (3.0–59) 14.1 (3.0–59) 16.0 (4.0–45) 11.6 (3.0–59) 14.9 (3.5–80) 17.7 (3.5–80) 10.9 (4.0–23) p-value \(\:{7.43\cdot\:10}_{}^{-9}\) Dermal Mitoses (count) 417 245 (59%) 172 (41%) 332 195 (59%) 137 (41%) 85 50 (59%) 35 (41%) p-value \(\:{3.56\cdot\:10}_{}^{-12}\) Ulceration (count) 426 248 (58%) 178 (42%) 341 198 (58%) 143 (41%) 85 50 (59%) 35 (41%) p-value \(\:{4.45\cdot\:10}_{}^{-6}\) Regression (count) 404 240 (59%) 164 (41%) 322 192 (60%) 130 (40%) 82 48 (59%) 34 (41%) p-value \(\:{1.74\cdot\:10}_{}^{-4}\) Time (Mean, IQR) Time between diagnosis and detection of metastases - 4.5 months (0–8 years) - - 4.7 months (0–8 years) - - 3.5 months (0–5 years) - Follow up time - - 5.3 years (3-7.3) - - 5.3 years (3-7.3) - - 5.2 years (3.5–6.8) Statistical Analysis The Mann-Whitney U test was conducted to compare each parameter between the metastatic and non-metastatic groups. The test was applied separately for each parameter, providing p-values that indicate whether the observed differences between the two groups were statistically significant. The test evaluates whether the distribution of values for a given parameter differs between the metastatic and non-metastatic groups; the p-values are presented in Table 2 . Image Feature Extraction The WSIs were tiled into 224 by 224 patches at 10X magnification using OpenSlide 34 . There was no overlap between the patches and only patches with at least 15% tissue were kept for further analysis. In total, 2.3 million patches were generated for the training set and 0.5 million patches were generated for the test set. After processing WSIs into patches, features were extracted using the whole-slide model Prov GigaPath 35 . Prov-GigaPath is a foundation model designed for analyzing gigapixel pathology slides by extracting slide-level embeddings for diverse clinical applications. It uses a two-stage approach with a tile encoder, pretrained using DINOv2 10 , to capture local features from image tiles, and a slide encoder, leveraging masked autoencoder pretraining with LongNet 36 , to model global features across the entire slide. Prov-GigaPath was pretrained on the dataset Prov-Path comprising over 1.38 billion tiles from 171,189 pathology slides, representing 31 tissue types and data from over 30,000 patients. Clinical Feature Extraction Text sentences were generated based on the tabular data presented in Table 2 . Each sentence followed a consistent structure, generated from three templates: “Whole slide image of malignant melanoma, has a breslow thickness of * mm and a diameter of * mm, shows *, *, *” “Whole slide image of malignant melanoma, has a breslow thickness of * mm, shows *, *, *” “Whole slide image of malignant melanoma, has a diameter of * mm, shows *, *, *” Here, * represents values extracted from the tabular data. For example, one generated sentence reads: "Whole slide image of malignant melanoma, has a Breslow thickness of 1.2 mm and a diameter of 13.0 mm, shows mitotic activity, regression, and ulceration." Text embeddings were then generated for each sentence using the pre-trained large language model BiomedBERT 37 . In cases where tabular data were missing, the corresponding information was simply omitted from the sentence, as BiomedBERT will generate embeddings of a fixed size regardless of input length. In the training dataset, only 306 WSIs had a complete set of histopathological features (184 in the metastatic group and 122 in the non-metastatic group). In the test dataset, 77 WSIs had a complete set of histopathological features (46 in the metastatic group and 31 in the non-metastatic group). Models MultiTrans MultiTrans incorporates a trainable layer that projects the text embeddings into Queries, while the image embeddings are projected into Keys and Values. These representations are then processed through Multi-Head Attention, producing attended features that are aggregated and passed through a final MLP layer for classification. Training the models Each model was trained on 341 samples using five-fold cross validation, with 80% of the data allocated for training and 20% for validation. A final model was obtained through majority voting across the five cross-validation models, and its performance was evaluated on an independent holdout test set consisting of 85 samples. The same hyperparameters were used for all models, with four attention heads in the Multi-Head Attention mechanism for both TransMIL and MultiTrans. Training was conducted with a learning rate of 5E − 5 , for 50 epochs, with early stopping applied after 10 epochs of no improvement. Optimization was performed using the Adam optimizer with a weight decay of 10E − 3. One bag (batch) was fed to the model at a time. Training was performed on a single GPU on a DGX A100 system and, on average, converged within 10 minutes for both TransMIL and MultiTrans. Declarations Acknowledgements The study was financed by grants from the Swedish state under the agreement between the Swedish government and the county councils, the ALF-agreement (Grant ALFGBG-991541), by Cancerfonden (Grant 23 2876 Pj), by Hudfonden and by Finnish Dermatopathologist foundation. Authors' contributions Conception and design: NN, FD, IP, IH, FY. Development of methodology: FD, FY, IP, IH, NN. Acquisition of data: NN, IS, JJ, OD. Analysis and interpretation of data: NN, FD, IP, IH, IS. Writing, reviewing, and revision of the manuscript: NN, FD, FY, IP, IH , IS, JJ, OD. Study supervision: NN, IP, IH, FY. Acquisition of funding: NN Data availability statement The datasets generated and/or analysed during the current study will be available at the melanoma dataset via AIDA (data transfer ongoing)https://datahub.aida.scilifelab.se/datasets/ The code will be published on GitHub (Code review ongoing). Additional Information (including a Competing Interests Statement) There are no conflicts of interest. Ethics approval and consent to participate The study was approved by the regional ethical committee of Gothenburg (Dnr 2023-06786-02). Since all the material was anonymized, consent to participate was waived. References Gershenwald, J. et al. Melanoma staging: Evidence-based changes in the american joint committee on cancer eighth edition cancer staging manual. CA Cancer J Clin 67 , 472–492 (2017). Thompson, J. & Williams, G. When does a melanoma metastasize? implications for management. Oncotarget 15, 374–378, DOI: 10.18632/oncotarget.28591 (2024). Ossowski, L. & Aguirre-Ghiso, J. Dormancy of metastatic melanoma. Pigment. Cell Melanoma Res 23 , 41–56 (2010). Adler, N., Haydon, A., McLean, C., Kelly, J. & Mar, V. Metastatic pathways in patients with cutaneous melanoma. Pigment. Cell Melanoma Res 30 , 13–27 (2017). Huang, H., Fu, Z., Ji, J., Huang, J. & Long, X. Predictive values of pathological and clinical risk factors for positivity of sentinel lymph node biopsy in thin melanoma: A systematic review and meta-analysis. Front Oncol 12, 817510, DOI: 10.3389/fonc.2022.817510 (2022). Balch, C. et al. Prognostic factors analysis of 17,600 melanoma patients: validation of the american joint committee on cancer melanoma staging system. J Clin Oncol 19, 3622–3634, DOI: 10.1200/JCO.2001.19.16.3622 (2001). Dillekås, H., Rogers, M. & Straume, O. Are 90 Cancer Med 8, 5574–5576 (2019). Campanella, G. et al. A clinical benchmark of public self-supervised pathology foundation models (2024). 2407.06508. Zhang, H. et al. Dino: Detr with improved denoising anchor boxes for end-to-end object detection (2022). 2203.03605. Oquab, M. et al. Dinov2: Learning robust visual features without supervision (2024). 2304.07193. Acs, B., Rantalainen, M. & Hartman, J. Artificial intelligence as the next step towards precision pathology. J Intern Med 288, 62–68 (2020). Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. (2021). 2010.11929. Shao, Z. et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Adv. Neural Inf. Process. Syst. 34, 2136–2147 (2021). Zeid, M. A.-E., El-Bahnasy, K. & Abo-Youssef, S. E. Multiclass colorectal cancer histology images classification using vision transformers. In 2021 Tenth International Conference on Intelligent Computing and Information Systems (ICICIS), 224–230, DOI: 10.1109/ICICIS52592.2021.9694125 (2021). Yacob, S. J. V. K. e. a., F. Weakly supervised detection and classification of basal cell carcinoma using graph-transformer on whole slide images. Sci. Reports DOI: https://doi.org/10.1038/s41598-023-33863-z (2023). Waqas, M., Ahmed, S. U., Tahir, M. A., Wu, J. & Qureshi, R. Exploring multiple instance learning (mil): A brief survey. Expert. Syst. with Appl. 250, 123893, DOI: https://doi.org/10.1016/j.eswa.2024.123893 (2024). De Logu, F. et al. Recognition of cutaneous melanoma on digitized histopathological slides via artificial intelligence algorithm. Front. Oncol. 10, 1559 (2020). Published online 2020 Aug 20. Hekler, A. et al. Pathologist-level classification of histopathological melanoma images with deep neural networks. Eur. J. Cancer 115, 79–83, DOI: 10.1016/j.ejca.2019.04.02 (2019). Sun, D. et al. Artificial intelligence-based pathological application to predict regional lymph node metastasis in papillary thyroid cancer. Curr. Probl. Cancer 53, 101150, DOI: https://doi.org/10.1016/j.currproblcancer.2024.101150 (2024). Tan, L. H. Y. J. e. a., L. Colorectal cancer lymph node metastasis prediction with weakly supervised transformer-based multi-instance learning., DOI: https://doi.org/10.1007/s11517-023-02799-x (2023). Lu, W. D. F. K. C. T. Y. C. R. J. B. M. . M. F., M. Y. Data-efficient and weakly supervised computational pathology on whole-slide images., DOI: https://doi.org/10.1038/s41551-020-00682-w (2021). Jansen, P. et al. Deep learning detection of melanoma metastases in lymph nodes. Eur. J. Cancer 188, 161–170, DOI: 10.1016/j.ejca.2023.04.023 (2023). Siarov, J. et al. Deep learning model shows pathologist-level detection of sentinel node metastasis of melanoma and intra-nodal nevi on whole slide images. Front. Medicine (Lausanne) 11, 1418013, DOI: 10.3389/fmed.2024.1418013 (2024). Published online 2024 Aug 22. Brinker, T. J. et al. Deep learning approach to predict sentinel lymph node status directly from routine histology of primary melanoma tumours. Eur. J. Cancer 154, 227–234, DOI: 10.1016/j.ejca.2021.05.026 (2021). Knuutila, R. P. K. A. T. M. T. L. N. L. . K. V. M., J. S. Identification of metastatic primary cutaneous squamous cell carcinoma utilizing artificial intelligence analysis of whole slide images, DOI: https://doi.org/10.1038/s41598-022-13696-y (2022). Kulkarni, P. M. et al. Deep learning based on standard he images of primary melanoma tumors identifies patients at risk for visceral recurrence and death. Clin. Cancer Res. 26, 1126–1134, DOI: 10.1158/1078-0432.CCR-19-1573 (2020). Shao, Z. et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classication. CoRR abs/2106.00908 (2021). 2106.00908. Bankhead, P., Loughrey, M. B., Fernandez, J. A. et al. Qupath: Open source software for digital pathology image analysis. Sci. Reports 7, 16878, DOI: 10.1038/s41598-017-17204-5 (2017). Meier, F. et al. Metastatic pathways and time courses in the orderly progression of cutaneous melanoma. Br J Dermatol 147, 62–70, DOI: 10.1046/j.1365-2133.2002.04867.x (2002). Leiter, U., Meier, F., Schittek, B. & Garbe, C. The natural course of cutaneous melanoma. J Surg Oncol 86 , 172–178, DOI: 10.1002/jso.20079 (2004). Mervic, L. Time course and pattern of metastasis of cutaneous melanoma differ between men and women. PLoS One 7 , e32955, DOI: 10.1371/journal.pone.0032955 (2012). Schmid-Wendtner, M. et al. Late metastases of cutaneous melanoma: an analysis of 31 patients. J Am Acad Dermatol 43 , 605–609, DOI: 10.1067/mjd.2000.107234 (2000). Zhai, X., Kolesnikov, A., Houlsby, N. & Beyer, L. Scaling vision transformers (2022). 2106.04560. Goode A, H. J. J. D. S. M., Gilbert B. Openslide: A vendor-neutral software foundation for digital pathology. J Pathol Inf. DOI: doi:10.4103/2153-3539.119005 (2013). Xu, H. et al. A whole-slide foundation model for digital pathology from real-world data. Nature (2024). Ding, J. et al. Longnet: Scaling transformers to 1,000,000,000 tokens (2023). 2307.02486. Gu, Y. et al. Domain-specific language model pretraining for biomedical natural language processing (2020). arXiv: 2007.15779. Vaswani, A. et al. Attention is all you need (2023). 1706.03762. Additional Declarations No competing interests reported. Supplementary Files Supplementarymaterial.docx Cite Share Download PDF Status: Published Journal Publication published 01 Apr, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 16 Dec, 2025 Reviews received at journal 25 Nov, 2025 Reviewers agreed at journal 24 Nov, 2025 Reviews received at journal 27 Sep, 2025 Reviewers agreed at journal 13 Sep, 2025 Reviewers agreed at journal 27 Jul, 2025 Reviewers agreed at journal 06 Apr, 2025 Reviewers invited by journal 06 Apr, 2025 Editor assigned by journal 06 Apr, 2025 Editor invited by journal 04 Apr, 2025 Submission checks completed at journal 02 Apr, 2025 First submitted to journal 24 Mar, 2025 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-6297243","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":433313192,"identity":"3cd0e523-e200-4ca2-8e71-2f7ae62d3d03","order_by":0,"name":"Filip Dahlén","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7ElEQVRIie3RsQqCQBjA8e84OJcj1wuHXsEIitDwQVqMIJeERocIQ2jqAXwM17bgA1t8i6BZCKKxT7Il0GpruD843HE/vE8BdLo/zEh4fATgYBu0KunpVKu2JLKa0DmWAijxkRxZDC/C5VfEYFtcgdMbcZ6f3Qg3wkjQhrU7bSScxZhC0D8kIhgsC1RC5nMf8kUYNxCvIhKQZSiHVrgjopYDui02ElkTL0PzZo1/ITN6i7BYTfyPJLWDeYZi2N0XQXdHs9h+yyzSRLyuImeSnZKLukeOadIXU+XabSRV9FPet/yW40+i0+l0urYeHKdPVP7wIyYAAAAASUVORK5CYII=","orcid":"","institution":"University of Jyväskylä","correspondingAuthor":true,"prefix":"","firstName":"Filip","middleName":"","lastName":"Dahlén","suffix":""},{"id":433313193,"identity":"168491ca-03f2-40ae-9809-830b5331dbbd","order_by":1,"name":"Ivan Shujski","email":"","orcid":"","institution":"Norra Älvsborg Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ivan","middleName":"","lastName":"Shujski","suffix":""},{"id":433313194,"identity":"b51edade-407e-4a22-b339-3b6ec09d4152","order_by":2,"name":"Filmon Yacob","email":"","orcid":"","institution":"Ekkono Solutions","correspondingAuthor":false,"prefix":"","firstName":"Filmon","middleName":"","lastName":"Yacob","suffix":""},{"id":433313195,"identity":"2569168a-d0aa-45e9-a0d7-7761089ae2a6","order_by":3,"name":"Ida Häggström","email":"","orcid":"","institution":"Chalmers University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Ida","middleName":"","lastName":"Häggström","suffix":""},{"id":433313196,"identity":"6ba172b1-11a4-415b-941f-2c57e4ddf875","order_by":4,"name":"Jovana Jovanovic","email":"","orcid":"","institution":"Södra Älvsborg Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jovana","middleName":"","lastName":"Jovanovic","suffix":""},{"id":433313197,"identity":"06f51493-5737-414c-9cc1-8adc60ecce60","order_by":5,"name":"Olga Dudina","email":"","orcid":"","institution":"Norra Älvsborg Hospital","correspondingAuthor":false,"prefix":"","firstName":"Olga","middleName":"","lastName":"Dudina","suffix":""},{"id":433313199,"identity":"12ba7bac-cf20-414f-baf9-abcca24472b6","order_by":6,"name":"Ilkka Pölönen","email":"","orcid":"","institution":"University of Jyväskylä","correspondingAuthor":false,"prefix":"","firstName":"Ilkka","middleName":"","lastName":"Pölönen","suffix":""},{"id":433313200,"identity":"d04f5b33-97ab-4a1f-9ca9-8d450010f7d8","order_by":7,"name":"Noora Neittaanmäki","email":"","orcid":"","institution":"University of Gothenburg","correspondingAuthor":false,"prefix":"","firstName":"Noora","middleName":"","lastName":"Neittaanmäki","suffix":""}],"badges":[],"createdAt":"2025-03-24 16:23:26","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6297243/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6297243/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-45588-w","type":"published","date":"2026-04-01T15:59:27+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":79265309,"identity":"b3441ba2-2344-4b05-ab4a-340a7da121db","added_by":"auto","created_at":"2025-03-26 09:59:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":141084,"visible":true,"origin":"","legend":"\u003cp\u003eMean receiver operating characteristic curves with noted AUC of MultiTrans from the five-fold cross validation tested on the hold-out test set (n=85).\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6297243/v1/f2eb4a3b969f2228470fb9a9.png"},{"id":79265307,"identity":"a62b46d8-5073-46f0-aa21-af9b00478738","added_by":"auto","created_at":"2025-03-26 09:59:11","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":149598,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrix of MultiTrans, evaluated on the test set (n=85) at the Youden index.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6297243/v1/13d52090e56ec30ba3fb722f.png"},{"id":79265834,"identity":"185640d6-f259-43af-886e-7f306a37aa32","added_by":"auto","created_at":"2025-03-26 10:07:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2027147,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of class activation maps derived from attention weights generated by MultiTrans over the original WSI (\u003cstrong\u003ed\u003c/strong\u003e, \u003cstrong\u003ee\u003c/strong\u003e and \u003cstrong\u003ef\u003c/strong\u003e) and the corresponding H\u0026amp;E image (\u003cstrong\u003ea\u003c/strong\u003e, \u003cstrong\u003eb\u003c/strong\u003e and \u003cstrong\u003ec\u003c/strong\u003e), with the tumor area annotated. Sample \u003cstrong\u003ea, d\u003c/strong\u003e represents a non-metastatic primary melanoma with a diameter of 7 mm and a Breslow thickness of 1.7 mm. It shows ulceration, dermal mitosis, and no regression. Sample \u003cstrong\u003eb, e\u003c/strong\u003erepresents a metastatic primary melanoma with a diameter of 16 mm and a Breslow thickness of 11 mm. It shows ulceration, dermal mitoses, and no regression. Sample \u003cstrong\u003ec, f\u003c/strong\u003e represents a metastatic primary melanoma with a diameter of 10 mm and a Breslow thickness of 6.5 mm. It shows ulceration, dermal mitoses, and no regression.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6297243/v1/eeaf6e431cefd211c50324aa.png"},{"id":79265314,"identity":"ddc7747f-2869-4967-b10f-ee1fed84785b","added_by":"auto","created_at":"2025-03-26 09:59:11","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":409663,"visible":true,"origin":"","legend":"\u003cp\u003eMethod overview. The WSI is first tiled into patches, with each patch converted into a feature embedding using Prov-GigaPath. For each corresponding set of clinical data, a text sentence is constructed, and text embeddings are generated using BiomedBERT. The image embeddings and text embeddings are then projected into queries, keys, and values through three linear layers. Attention is computed between the image and text embeddings, and the aggregated output is classified using an MLP layer.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6297243/v1/2af6b57ca44b8216980cd28c.png"},{"id":106344460,"identity":"7c258676-ac07-4802-b052-9a5c042fefd3","added_by":"auto","created_at":"2026-04-07 16:14:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3388928,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6297243/v1/ac58fbf3-9a70-47cf-99e3-4152116bdb3f.pdf"},{"id":79265306,"identity":"8dcfbb08-f6b7-4001-915d-158fd2230d2f","added_by":"auto","created_at":"2025-03-26 09:59:11","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":122383,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarymaterial.docx","url":"https://assets-eu.researchsquare.com/files/rs-6297243/v1/977013153c67557f852da2a3.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Predicting metastatic potential of primary cutaneous melanomas utilizing weakly supervised vision language model","fulltext":[{"header":"Introduction","content":"\u003cp\u003eCutaneous melanoma is one of the deadliest skin cancers and a growing cause of death and morbidity worldwide. The prognosis for stage III and IV melanoma patients is significantly worse compared to patients in stage I or II without metastatic disease\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Thus, adjuvant treatments with both targeted therapies and immunotherapy are often considered for patients in stage III and IV\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. There is evidence indicating that metastases from a primary cutaneous melanoma often occur long before the melanoma is diagnosed\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. However, some melanomas metastases are detected early, while some melanomas stay in a \u0026ldquo;dormant\u0026rdquo; stage for years\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Melanoma metastasizes lymphogenously (to regional lymph nodes or as in-transit metastases), but also hematogenously with distant metastases\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThus, knowing if a melanoma is likely to metastasize is crucial for treatment and survival of melanoma patients. Primary tumors contain morphological markers that may indicate metastatic potential. At the time of a primary melanoma diagnosis, it is possible to use clinico-pathological parameters to predict the likelihood of sentinel lymph node metastases\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e and of eventual death from melanoma\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. The most informative prognostic features for survival are the Breslow thickness, ulceration and presence of dermal mitoses\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Current diagnostic practices primarily involve histopathological evaluation, which can be labor-intensive and subject to interobserver variability.\u003c/p\u003e \u003cp\u003eDespite the progress of self-supervised learning and foundation models in computer vision and natural language processing, their application in the medical domain remains in its early stages. One of the main reasons is the limited availability of publicly accessible data compared to other domains \u0026mdash; a challenge particularly pronounced in pathology due to the relatively slow adoption of digital pathology and the inevitable time constraints in collecting patient data. However, recent efforts have led to the public release of several foundation models pre-trained on large-scale clinical datasets. These developments have significantly advanced computational pathology research, by lowering the barriers for smaller research groups and accelerating the translation of AI-based methods into clinical practice. Campanella et al.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e made a clinical benchmark of public self-supervised pathology foundation models and measured the performance of the models on two types of downstream tasks, disease detection and computational biomarkers. In general, the models trained using DINO\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e and DINOv2\u003csup\u003e10\u003c/sup\u003e (SP21M, SP85M, UNI, Virchow, and Prov-GigaPath) achieved comparable performance and in biomarker prediction UNI and Prov-Gigapath were just as good or better than the other models with Prov-Gigapth performing slightly better than UNI. In biomarker prediction, an overall trend towards higher performance with larger models was also observed.\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThese emerging self-supervised models provide a promising starting point for further advancements in computational pathology. Alongside these developments, recent developments in computational pathology have enabled development of prognostic models based on digitized routine stained histopathological whole slide images (WSIs)\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e. Significant progress has been made in detecting and classifying cancers and identifying metastases in lymph nodes using machine learning models. In the domain of histological imaging, both Vision Transformers\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e and more traditional convolutional neural network (CNN) architectures have demonstrated significant success in a range of applications, such as detecting breast cancer metastases and classifying cancer subtypes in lung, kidney and colorectal tissues\u003csup\u003e\u003cspan additionalcitationids=\"CR14 CR15 CR16 CR17\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Furthermore, transformer based models, CNN models and MIL models have been particularly effective in detection of metastases\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan additionalcitationids=\"CR20 CR21 CR22\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. Despite these advancements, only a limited number of studies have focused on identifying metastatic potential in primary tumors including Brinker et al.\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e which predicted sentinel lymph node metastatic status directly from routine histology of primary melanoma tumors using a CNN-based approach. Similarly, Knuutila et al.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e investigated metastatic primary cutaneous squamous cell carcinoma instead using a residual neural network (RNN) architecture. In another study, Kulkarni et al.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e analyzed primary melanoma tumors to identify patients at risk for visceral recurrence and death using a CNN combined with an RNN architecture.\u003c/p\u003e \u003cp\u003eIn this study, we aimed to predict metastatic risk in primary melanomas by developing a multimodal transformer trained on both image and text embeddings referred to as MultiTrans. Furthermore, we aimed to conduct an ablation study to compare this model to TransMIL\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e trained solely with image embeddings and a MLP trained only with text embeddings referred to as BertMLP.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eMultiTrans achieved an AUC of 0.887, an accuracy of 0.871 with a sensitivity of 0.921 and a specificity of 0.794 when evaluated at the Youden index. The mean receiver operating characteristic (ROC) curve is visualized in Fig 1, the confusion matrix at the Youden index is visualized in Fig 2. To further interpret the model\u0026rsquo;s predictions, heatmaps were used to highlight regions in whole slide images (WSI) that contributed most to classification. Figure 3 displays tumor regions in three different samples correctly identified by MultiTrans.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAblation Study\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate the impact of different model architectures, we compared MultiTrans, TransMIL, and BertMLP on the hold-out test set.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eTransMIL\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe TransMIL as described by Shao et al.\u003csup\u003e27\u0026nbsp;\u003c/sup\u003ewas used. The TransMil model, given a bag of image embeddings, will perform multi-head self-attention\u003csup\u003e38\u0026nbsp;\u003c/sup\u003eon the input embeddings to capture the correlation between the different embeddings. Furthermore, the attended features are fed to a Pyramid Position Encoding Generator (PPEG) that will encode the spatial information before being fed to the last Transformer layer to aggregate the morphological information. Finally, the aggregated features are fed to a single MLP layer for classification.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eBertMLP\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eClassification of text embeddings was performed using a MLP consisting of three layers with ReLU activation functions in between.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe full performance metrics are summarized in Table 1. The mean ROC curve for each model is visualized in Fig S1, while confusion matrices evaluated at the Youden index are shown in Fig S2. MultiTrans achieved the highest accuracy (0.871) compared to both TransMIL (0.856) and BertMLP (0.729) when evaluated at the Youden index. Both MultiTrans and TransMIL demonstrated high AUC and sensitivity, indicating strong ability to identify metastatic cases with MultiTrans performing slightly better (0.921 vs 0.882), detecting more metastatic cases than TransMIL (47 TP vs 45 TP). However, TransMIL achieved higher specificity (0.824 vs 0.794) indicating it reduced false positives more effectively (6 FP vs 7 FP). The AUC for MultiTrans and TransMIL were similar with slightly higher values for MultiTrans (0.887 vs 0.883). MultiTrans and TransMIL significantly outperformed BertMLP across all metrics besides Specificity where BertMLP had higher value than MultiTrans (0.882 vs 0.794), meaning it produced fewer false positives than MultiTrans (4 FP vs 7 FP). MultiTrans minimized false negatives (FN = 4), which is crucial for ensuring metastatic cases are not missed. BertMLP had the highest false negatives (FN = 19), making it unreliable for detecting metastatic cases. These results suggest that WSIs alone provided sufficient information for accurate predictions, while the inclusion of histopathological features further enhances model performance.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eTable 1\u003c/strong\u003e:\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003ePerformance metrics for MultiTrans, TransMIL and BertMLP.\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"657\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 79px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003eSensitivity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003eSpecificity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003eTP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003eFP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003eFN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003eTN\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 79px;\"\u003e\n \u003cp\u003eMultiTrans\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.871\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.921\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e0.794\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.887\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e47\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e4\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e27\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 79px;\"\u003e\n \u003cp\u003eTransMIL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003e0.856\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e0.882\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e0.824\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003e0.883\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 79px;\"\u003e\n \u003cp\u003eBertMLP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003e0.729\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e0.627\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.882\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 65px;\"\u003e\n \u003cp\u003e0.802\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 73px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e30\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this paper, we predicted metastatic potential in primary melanomas by comparing a TransMIL trained on histopathological image data, MultiTrans trained on image data together with clinical data, and a BertMLP trained on clinical data. The results show that TransMIL is able to detect early signs of metastatic potential in the primary tumors with relatively high accuracy. MultiTrans achieves a slightly higher AUC compared to TransMIL but not large enough to be significant. Both TransMIL and MultiTrans outperform BertMLP. The results highlight the impact of using image data for this prediction task.\u003c/p\u003e \u003cp\u003eBrinker et al.\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e predicted the sentinel lymph node metastatic status directly from routine histology of primary melanoma tumors, achieving an AUC of 0.618 when using WSIs alone, 0.616 when using only histopathological features, and a slightly lower AUC of 0.613 when combining WSIs and histopathological features. Similarly, Knuutila et al.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e investigated metastatic primary cutaneous squamous cell carcinoma using a residual neural network architecture, reporting an AUC of 0.747 when analyzing WSIs alone, 0.804 when using only histopathological features, and an AUC of 0.917 when developing a risk factor model (RFM) consisting of AI-based WSI predictions, Clark\u0026rsquo;s level 5, and tumor diameter\u0026thinsp;\u003cem\u003e\u0026ge;\u003c/em\u003e\u0026thinsp;40 mm as risk factors. In another study, Kulkarni et al.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e analyzed primary melanoma tumors to identify patients at risk of visceral recurrence and death. They employed a CNN combined with a RNN architecture, achieving an AUC of 0.905 and 0.880 in two independent validation sets.\u003c/p\u003e \u003cp\u003eIn the study by Brinker et al.\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, tumor regions were first identified and annotated as regions of interest for each WSI, followed by division into patches, which were processed through a CNN and subsequently classified using an MLP layer. Similarly, Knuutila et al.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e divided each WSI into patches, assigned binary tumor labels based on prior annotations, and further labeled them according to metadata as indicating rapid metastases or not. The patches were then classified using a ResNet-18 architecture. Kulkarni et al.\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e pre-processed the WSIs to isolate tumor regions using QuPath digital pathology software. In contrast to these approaches, which relied on detailed annotations or patch-level labeling, our method is weakly supervised, with no predefined annotations or patch-level labels. Instead, the transformer model learned to identify discriminative regions through self-attention over all patches within each WSI. Additionally, none of the previous studies utilized a foundation model for feature extraction as in our approach, where feature extraction is decoupled from the classification task.\u003c/p\u003e \u003cp\u003eThe findings from our study suggest that WSIs alone provided sufficient information for accurate predictions, while the inclusion of histopathological features further enhances model performance. These results align with the prior study by Knuutila et al.\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, which demonstrated improved predictive performance when combining image-based predictions with clinical risk factors. Our study also highlights the importance of utilizing a foundation model to extract relevant features from the image patches together with a transformer to direct attention towards the most relevant regions which is supported by our results being an AUC of 0.883 compared to 0.618 and 0.747\u003csup\u003e26\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eHowever, the observation that the combined use of WSIs and histopathological features yielded only a marginal improvement over WSIs alone may indicate that the applied data fusion strategy was suboptimal. The approach used in this study may not have effectively aligned or balanced the image embeddings with the histopathological features, potentially limiting the model\u0026rsquo;s overall performance. Future research should explore more advanced data fusion methods and the development of multimodal embeddings to better integrate different data sources and improve predictive performance.\u003c/p\u003e \u003cp\u003eEvidence suggests that melanoma metastases often occurred many months before a primary melanoma diagnosis is made\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e which makes it crucial to identify these aggressive tumors in an early stage. Metastases to regional lymph nodes via lymphatics is the most common form of spread, with around 50% of those who develop metastases having nodal disease as their first site of clinically-detected recurrence\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. However, metastases in a distant organ with no evidence of previous or current lymph node disease is seen in about 30% of those who develop metastatic melanoma\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e,\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e suggesting dissemination of malignant cells exclusively via the bloodstream. Lymph node metastases are often diagnosed earlier than metastases at distant sites, with a median interval of 16 months between primary diagnosis and the detection of nodal metastases in one study\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, while distant metastases tend to be detected a median of 25\u0026ndash;40 months after primary diagnosis\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. In our study of 426 tumors, 249 were metastatic, and of those 128 were in stage III showing lymph node or local in transit metastases while 121 showed distal metastases (stage IV). Despite the three year follow up and sentinel node biopsies conducted in all the patients in the non-metastatic group, some of the non-metastatic group may have developed undetected micrometastatic disease. Furthermore, occasionally melanomas show very late clinical appearance of metastases, sometimes more than 10 years after the primary melanoma was excised\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eOne limitation in our study is the dataset size. A dataset of 426 WSIs, while valuable, may not capture the full variability of tumor morphology and metastatic patterns present in a broader population. This limitation could potentially hinder the generalizability of the model to unseen data or rare cases. The dataset size plays a critical role in improving model performance\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. Furthermore, it has been shown that simultaneously scaling the dataset size and the model size can lead to significant performance gains. Another limitation is that we didn\u0026rsquo;t predict disease-specific survival. Additionally, not all WSI had a complete set of histopathological features which most possibly lowered the performance of the models. Also, every WSI had an identification number written on the glass when scanned that could potentially influence the performance, but according to the attention maps, no attention was laid on these areas. We evaluated two sets of text sentences for BertMLP: one including gender and age information and one without, which represents the current approach. The results showed that incorporating gender and age led to lower performance compared to the current approach, which contrasts with the findings of Mervic et al.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. One possible explanation for this discrepancy is the imbalance in the dataset, where males outnumber females, potentially introducing bias. However, this warrants further investigation\u003c/p\u003e \u003cp\u003eTo conclude, TransMIL trained solely on image data was able to detect early signs of metastatic potential in the primary tumors with high accuracy outperforming the BertMLP trained with the histopathological parameters which are the current prognostic standard. MultiTrans combining the text embeddings of histopathological parameters and image data achieved a slightly higher AUC compared to TransMIL and somewhat higher accuracy when evaluated at the Youden index. The results highlight the impact of using image data for this prediction task and the potential in combining image data together with histopathological data and demonstrate high accuracy for early recognition of melanomas with high metastatic potential. Predicting early signs of metastatic potential from primary tumors could enable early targeted treatments for this patient group.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eAn overview of the method is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eDataset\u003c/h2\u003e \u003cp\u003eThe data was retrospectively collected at the Sahlgrenska University Hospital between 2016\u0026ndash;2023. Inclusion criteria were a primary cutaneous malignant melanoma stage I-IV with information about possible metastases and in case of non-metastatic tumor, a negative sentinel node examination and a minimum of 3 years follow-up time. The exclusion criteria were patient cases with metastatic malignant melanoma having more than one primary malignant melanoma tumor to avoid uncertainty of which was the primary tumor that metastasized. One glass slide per tumor harboring the largest Breslow thickness was collected and scanned unidentified using a scanner NanoZoomer S360 Hamamatsu at 40X magnification. The complete dataset consisted of 426 WSIs representing 426 primary melanomas (249 metastatic and 177 non-metastatic) from 425 patients (one patient had two non-metastatic melanomas), detailed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The size of the WSI ranged from 1.4GB up to 5.8GB with dimensions between (157 440 x 55 296) and up to (215 040 x 109824) pixels. The total size of the dataset was around 1.4TB.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDistribution of clinical and histopathological features across metastatic and non-metastatic classes. Breslow thickness and diameter are presented as mean (min - max) values in millimeters. Mitoses, ulceration, and regression are reported as the fraction of occurrences (present/not present) within each class.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eAll included\u003c/p\u003e \u003cp\u003eTotal Met Non-met\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eTraining and validation set\u003c/p\u003e \u003cp\u003eTotal Met Non-met\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e \u003cp\u003eHold-out test set\u003c/p\u003e \u003cp\u003eTotal Met Non-met\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSlides\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWSI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e426\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e249\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e177\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e341\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e198\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e143\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e34\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePatches\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2 838 864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 659 336\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 179 528\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2 272 424\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1 319 472\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e952 952\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e566 440\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e339 864\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e226 576\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eGender\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e231\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e138\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e182\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e107\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e18\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e195\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e111\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e159\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAge (Mean, IQR)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e66\u003c/p\u003e \u003cp\u003e(29\u0026ndash;92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e66\u003c/p\u003e \u003cp\u003e(32\u0026ndash;92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e65\u003c/p\u003e \u003cp\u003e(29\u0026ndash;90)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e65\u003c/p\u003e \u003cp\u003e(29\u0026ndash;92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e65\u003c/p\u003e \u003cp\u003e(32\u0026ndash;92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e65\u003c/p\u003e \u003cp\u003e(29\u0026ndash;89)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e68\u003c/p\u003e \u003cp\u003e(42\u0026ndash;90)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e68\u003c/p\u003e \u003cp\u003e(43\u0026ndash;89)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e67\u003c/p\u003e \u003cp\u003e(42\u0026ndash;90)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e65\u003c/p\u003e \u003cp\u003e(17\u0026ndash;94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e67\u003c/p\u003e \u003cp\u003e(21\u0026ndash;94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e63\u003c/p\u003e \u003cp\u003e(17\u0026ndash;92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e65\u003c/p\u003e \u003cp\u003e(17\u0026ndash;94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e67\u003c/p\u003e \u003cp\u003e(21\u0026ndash;94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e63\u003c/p\u003e \u003cp\u003e(17\u0026ndash;92)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e64\u003c/p\u003e \u003cp\u003e(27\u0026ndash;94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e64\u003c/p\u003e \u003cp\u003e(27\u0026ndash;94)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e65\u003c/p\u003e \u003cp\u003e(32\u0026ndash;86)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eStage\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIII\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e104\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e121\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ehistopathological features\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBreslow (mean, std, p-value)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3.7\u003c/p\u003e \u003cp\u003e(0.3\u0026ndash;35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.9\u003c/p\u003e \u003cp\u003e(0.5\u0026ndash;35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1.9\u003c/p\u003e \u003cp\u003e(0.3\u0026ndash;9.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.5\u003c/p\u003e \u003cp\u003e(0.5\u0026ndash;20)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.6\u003c/p\u003e \u003cp\u003e(0.5\u0026ndash;20)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e1.9\u003c/p\u003e \u003cp\u003e(0.8\u0026ndash;9.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e4.9\u003c/p\u003e \u003cp\u003e(0.3\u0026ndash;35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e6.3\u003c/p\u003e \u003cp\u003e(1.1\u0026ndash;35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e1.8\u003c/p\u003e \u003cp\u003e(0.3\u0026ndash;0.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{1.70\\cdot\\:10}_{}^{-23}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiameter\u003c/p\u003e \u003cp\u003e(mean, std)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.3\u003c/p\u003e \u003cp\u003e(3.0\u0026ndash;80.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e16.3\u003c/p\u003e \u003cp\u003e(3.5\u0026ndash;80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e11.6\u003c/p\u003e \u003cp\u003e(3.0\u0026ndash;59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e14.1\u003c/p\u003e \u003cp\u003e(3.0\u0026ndash;59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e16.0\u003c/p\u003e \u003cp\u003e(4.0\u0026ndash;45)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e11.6\u003c/p\u003e \u003cp\u003e(3.0\u0026ndash;59)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e14.9\u003c/p\u003e \u003cp\u003e(3.5\u0026ndash;80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e17.7\u003c/p\u003e \u003cp\u003e(3.5\u0026ndash;80)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e10.9\u003c/p\u003e \u003cp\u003e(4.0\u0026ndash;23)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{7.43\\cdot\\:10}_{}^{-9}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDermal Mitoses (count)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e417\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e245\u003c/p\u003e \u003cp\u003e(59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e172\u003c/p\u003e \u003cp\u003e(41%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e332\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e195\u003c/p\u003e \u003cp\u003e(59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e137\u003c/p\u003e \u003cp\u003e(41%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e50\u003c/p\u003e \u003cp\u003e(59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e35\u003c/p\u003e \u003cp\u003e(41%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{3.56\\cdot\\:10}_{}^{-12}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUlceration\u003c/p\u003e \u003cp\u003e(count)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e426\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e248\u003c/p\u003e \u003cp\u003e(58%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e178\u003c/p\u003e \u003cp\u003e(42%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e341\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e198\u003c/p\u003e \u003cp\u003e(58%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e143\u003c/p\u003e \u003cp\u003e(41%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e50\u003c/p\u003e \u003cp\u003e(59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e35\u003c/p\u003e \u003cp\u003e(41%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{4.45\\cdot\\:10}_{}^{-6}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegression\u003c/p\u003e \u003cp\u003e(count)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e404\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e240\u003c/p\u003e \u003cp\u003e(59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e164\u003c/p\u003e \u003cp\u003e(41%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e322\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e192\u003c/p\u003e \u003cp\u003e(60%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e130\u003c/p\u003e \u003cp\u003e(40%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e48\u003c/p\u003e \u003cp\u003e(59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e34\u003c/p\u003e \u003cp\u003e(41%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{1.74\\cdot\\:10}_{}^{-4}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTime (Mean, IQR)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTime between diagnosis and detection of metastases\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.5 months\u003c/p\u003e \u003cp\u003e(0\u0026ndash;8 years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4.7 months\u003c/p\u003e \u003cp\u003e(0\u0026ndash;8 years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e3.5 months\u003c/p\u003e \u003cp\u003e(0\u0026ndash;5 years)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFollow up time\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5.3 years\u003c/p\u003e \u003cp\u003e(3-7.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5.3 years\u003c/p\u003e \u003cp\u003e(3-7.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e-\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c10\"\u003e \u003cp\u003e5.2 years\u003c/p\u003e \u003cp\u003e(3.5\u0026ndash;6.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eThe Mann-Whitney U test was conducted to compare each parameter between the metastatic and non-metastatic groups. The test was applied separately for each parameter, providing p-values that indicate whether the observed differences between the two groups were statistically significant. The test evaluates whether the distribution of values for a given parameter differs between the metastatic and non-metastatic groups; the p-values are presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eImage Feature Extraction\u003c/h3\u003e\n\u003cp\u003eThe WSIs were tiled into 224 by 224 patches at 10X magnification using OpenSlide\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e. There was no overlap between the patches and only patches with at least 15% tissue were kept for further analysis. In total, 2.3\u0026nbsp;million patches were generated for the training set and 0.5\u0026nbsp;million patches were generated for the test set. After processing WSIs into patches, features were extracted using the whole-slide model Prov GigaPath\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. Prov-GigaPath is a foundation model designed for analyzing gigapixel pathology slides by extracting slide-level embeddings for diverse clinical applications. It uses a two-stage approach with a tile encoder, pretrained using DINOv2\u003csup\u003e10\u003c/sup\u003e, to capture local features from image tiles, and a slide encoder, leveraging masked autoencoder pretraining with LongNet\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e, to model global features across the entire slide. Prov-GigaPath was pretrained on the dataset Prov-Path comprising over 1.38\u0026nbsp;billion tiles from 171,189 pathology slides, representing 31 tissue types and data from over 30,000 patients.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eClinical Feature Extraction\u003c/h2\u003e \u003cp\u003eText sentences were generated based on the tabular data presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Each sentence followed a consistent structure, generated from three templates:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e\u0026ldquo;Whole slide image of malignant melanoma, has a breslow thickness of * mm and a diameter of * mm, shows *, *, *\u0026rdquo;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e\u0026ldquo;Whole slide image of malignant melanoma, has a breslow thickness of * mm, shows *, *, *\u0026rdquo;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e\u0026ldquo;Whole slide image of malignant melanoma, has a diameter of * mm, shows *, *, *\u0026rdquo;\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eHere, * represents values extracted from the tabular data. For example, one generated sentence reads:\u003c/p\u003e \u003cp\u003e\"Whole slide image of malignant melanoma, has a Breslow thickness of 1.2 mm and a diameter of 13.0 mm, shows mitotic activity, regression, and ulceration.\"\u003c/p\u003e \u003cp\u003eText embeddings were then generated for each sentence using the pre-trained large language model BiomedBERT\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. In cases where tabular data were missing, the corresponding information was simply omitted from the sentence, as BiomedBERT will generate embeddings of a fixed size regardless of input length.\u003c/p\u003e \u003cp\u003eIn the training dataset, only 306 WSIs had a complete set of histopathological features (184 in the metastatic group and 122 in the non-metastatic group). In the test dataset, 77 WSIs had a complete set of histopathological features (46 in the metastatic group and 31 in the non-metastatic group).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eModels\u003c/h2\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003eMultiTrans\u003c/h2\u003e \u003cp\u003eMultiTrans incorporates a trainable layer that projects the text embeddings into Queries, while the image embeddings are projected into Keys and Values. These representations are then processed through Multi-Head Attention, producing attended features that are aggregated and passed through a final MLP layer for classification.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eTraining the models\u003c/h2\u003e \u003cp\u003eEach model was trained on 341 samples using five-fold cross validation, with 80% of the data allocated for training and 20% for validation. A final model was obtained through majority voting across the five cross-validation models, and its performance was evaluated on an independent holdout test set consisting of 85 samples. The same hyperparameters were used for all models, with four attention heads in the Multi-Head Attention mechanism for both TransMIL and MultiTrans. Training was conducted with a learning rate of 5E\u003csup\u003e\u003cem\u003e\u0026minus;\u003c/em\u003e\u0026thinsp;5\u003c/sup\u003e, for 50 epochs, with early stopping applied after 10 epochs of no improvement. Optimization was performed using the Adam optimizer with a weight decay of 10E\u0026thinsp;\u003cem\u003e\u0026minus;\u003c/em\u003e\u0026thinsp;3. One bag (batch) was fed to the model at a time. Training was performed on a single GPU on a DGX A100 system and, on average, converged within 10 minutes for both TransMIL and MultiTrans.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was financed by grants from the Swedish state under the agreement between the Swedish government and the county councils, the ALF-agreement (Grant ALFGBG-991541), by Cancerfonden (Grant 23 2876 Pj), by Hudfonden and by Finnish Dermatopathologist foundation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConception and design: NN, FD, IP, IH, FY. \u0026nbsp; Development of methodology: FD, FY, IP, IH, NN. Acquisition of data: NN, IS, JJ, OD. Analysis and interpretation of data: NN, FD, IP, IH, IS. Writing, reviewing, and revision of the manuscript: NN, FD, FY, IP, IH , \u0026nbsp;IS, \u0026nbsp; JJ, OD. \u0026nbsp;Study supervision: NN, IP, IH, FY. \u0026nbsp;Acquisition of funding: NN\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and/or analysed during the current study will be available at the melanoma dataset via AIDA (data transfer ongoing)https://datahub.aida.scilifelab.se/datasets/\u003c/p\u003e\n\u003cp\u003eThe code will be published on GitHub (Code review ongoing).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAdditional Information\u003c/strong\u003e (including a Competing Interests Statement)\u003c/p\u003e\n\u003cp\u003eThere are no conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was approved by the regional ethical committee of Gothenburg (Dnr 2023-06786-02). Since all the material was anonymized, consent to participate was waived.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eGershenwald, J. \u003cem\u003eet al.\u0026nbsp;\u003c/em\u003eMelanoma staging: Evidence-based changes in the american joint committee on cancer eighth edition cancer staging manual. \u003cem\u003eCA Cancer J Clin\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e67\u003c/strong\u003e, 472\u0026ndash;492 (2017).\u003c/li\u003e\n \u003cli\u003eThompson, J. \u0026amp; Williams, G. When does a melanoma metastasize? implications for management. \u003cem\u003eOncotarget\u0026nbsp;\u003c/em\u003e15, 374\u0026ndash;378, DOI: 10.18632/oncotarget.28591 (2024).\u003c/li\u003e\n \u003cli\u003eOssowski, L. \u0026amp; Aguirre-Ghiso, J. Dormancy of metastatic melanoma. \u003cem\u003ePigment. Cell Melanoma Res\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 41\u0026ndash;56 (2010).\u003c/li\u003e\n \u003cli\u003eAdler, N., Haydon, A., McLean, C., Kelly, J. \u0026amp; Mar, V. Metastatic pathways in patients with cutaneous melanoma. \u003cem\u003ePigment. Cell Melanoma Res\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e30\u003c/strong\u003e, 13\u0026ndash;27 (2017).\u003c/li\u003e\n \u003cli\u003eHuang, H., Fu, Z., Ji, J., Huang, J. \u0026amp; Long, X. Predictive values of pathological and clinical risk factors for positivity of sentinel lymph node biopsy in thin melanoma: A systematic review and meta-analysis. Front Oncol 12, 817510, DOI: 10.3389/fonc.2022.817510 (2022).\u003c/li\u003e\n \u003cli\u003eBalch, C. et al. Prognostic factors analysis of 17,600 melanoma patients: validation of the american joint committee on cancer melanoma staging system. J Clin Oncol 19, 3622\u0026ndash;3634, DOI: 10.1200/JCO.2001.19.16.3622 (2001).\u003c/li\u003e\n \u003cli\u003eDillek\u0026aring;s, H., Rogers, M. \u0026amp; Straume, O. Are 90 Cancer Med 8, 5574\u0026ndash;5576 (2019).\u003c/li\u003e\n \u003cli\u003eCampanella, G. et al. A clinical benchmark of public self-supervised pathology foundation models (2024). 2407.06508.\u003c/li\u003e\n \u003cli\u003eZhang, H. et al. Dino: Detr with improved denoising anchor boxes for end-to-end object detection (2022). 2203.03605.\u003c/li\u003e\n \u003cli\u003eOquab, M. et al. Dinov2: Learning robust visual features without supervision (2024). 2304.07193.\u003c/li\u003e\n \u003cli\u003eAcs, B., Rantalainen, M. \u0026amp; Hartman, J. Artificial intelligence as the next step towards precision pathology. J Intern Med 288, 62\u0026ndash;68 (2020).\u003c/li\u003e\n \u003cli\u003eDosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. (2021). 2010.11929.\u003c/li\u003e\n \u003cli\u003eShao, Z. et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification. Adv. Neural Inf. Process. Syst. 34, 2136\u0026ndash;2147 (2021).\u003c/li\u003e\n \u003cli\u003eZeid, M. A.-E., El-Bahnasy, K. \u0026amp; Abo-Youssef, S. E. Multiclass colorectal cancer histology images classification using vision transformers. In 2021 Tenth International Conference on Intelligent Computing and Information Systems (ICICIS), 224\u0026ndash;230, DOI: 10.1109/ICICIS52592.2021.9694125 (2021).\u003c/li\u003e\n \u003cli\u003eYacob, S. J. V. K. e. a., F. Weakly supervised detection and classification of basal cell carcinoma using graph-transformer on whole slide images. Sci. Reports DOI: https://doi.org/10.1038/s41598-023-33863-z (2023).\u003c/li\u003e\n \u003cli\u003eWaqas, M., Ahmed, S. U., Tahir, M. A., Wu, J. \u0026amp; Qureshi, R. Exploring multiple instance learning (mil): A brief survey. Expert. Syst. with Appl. 250, 123893, DOI: https://doi.org/10.1016/j.eswa.2024.123893 (2024).\u003c/li\u003e\n \u003cli\u003eDe Logu, F. et al. Recognition of cutaneous melanoma on digitized histopathological slides via artificial intelligence algorithm. Front. Oncol. 10, 1559 (2020). Published online 2020 Aug 20.\u003c/li\u003e\n \u003cli\u003eHekler, A. et al. Pathologist-level classification of histopathological melanoma images with deep neural networks. Eur. J. Cancer 115, 79\u0026ndash;83, DOI: 10.1016/j.ejca.2019.04.02 (2019).\u003c/li\u003e\n \u003cli\u003eSun, D. et al. Artificial intelligence-based pathological application to predict regional lymph node metastasis in papillary thyroid cancer. Curr. Probl. Cancer 53, 101150, DOI: https://doi.org/10.1016/j.currproblcancer.2024.101150 (2024).\u003c/li\u003e\n \u003cli\u003eTan, L. H. Y. J. e. a., L. Colorectal cancer lymph node metastasis prediction with weakly supervised transformer-based multi-instance learning., DOI: https://doi.org/10.1007/s11517-023-02799-x (2023).\u003c/li\u003e\n \u003cli\u003eLu, W. D. F. K. C. T. Y. C. R. J. B. M. . M. F., M. Y. Data-efficient and weakly supervised computational pathology on whole-slide images., DOI: https://doi.org/10.1038/s41551-020-00682-w (2021).\u003c/li\u003e\n \u003cli\u003eJansen, P. et al. Deep learning detection of melanoma metastases in lymph nodes. Eur. J. Cancer 188, 161\u0026ndash;170, DOI: 10.1016/j.ejca.2023.04.023 (2023).\u003c/li\u003e\n \u003cli\u003eSiarov, J. et al. Deep learning model shows pathologist-level detection of sentinel node metastasis of melanoma and intra-nodal nevi on whole slide images. Front. Medicine (Lausanne) 11, 1418013, DOI: 10.3389/fmed.2024.1418013 (2024). Published online 2024 Aug 22.\u003c/li\u003e\n \u003cli\u003eBrinker, T. J. et al. Deep learning approach to predict sentinel lymph node status directly from routine histology of primary melanoma tumours. Eur. J. Cancer 154, 227\u0026ndash;234, DOI: 10.1016/j.ejca.2021.05.026 (2021).\u003c/li\u003e\n \u003cli\u003eKnuutila, R. P. K. A. T. M. T. L. N. L. . K. V. M., J. S. Identification of metastatic primary cutaneous squamous cell carcinoma utilizing artificial intelligence analysis of whole slide images, DOI: https://doi.org/10.1038/s41598-022-13696-y (2022).\u003c/li\u003e\n \u003cli\u003eKulkarni, P. M. et al. Deep learning based on standard he images of primary melanoma tumors identifies patients at risk for visceral recurrence and death. Clin. Cancer Res. 26, 1126\u0026ndash;1134, DOI: 10.1158/1078-0432.CCR-19-1573 (2020).\u003c/li\u003e\n \u003cli\u003eShao, Z. et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classication. CoRR abs/2106.00908 (2021). 2106.00908.\u003c/li\u003e\n \u003cli\u003eBankhead, P., Loughrey, M. B., Fernandez, J. A. et al. Qupath: Open source software for digital pathology image analysis. Sci. Reports 7, 16878, DOI: 10.1038/s41598-017-17204-5 (2017).\u003c/li\u003e\n \u003cli\u003eMeier, F. et al. Metastatic pathways and time courses in the orderly progression of cutaneous melanoma. Br J Dermatol 147, 62\u0026ndash;70, DOI: 10.1046/j.1365-2133.2002.04867.x (2002).\u003c/li\u003e\n \u003cli\u003eLeiter, U., Meier, F., Schittek, B. \u0026amp; Garbe, C. The natural course of cutaneous melanoma. \u003cem\u003eJ Surg Oncol\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e86\u003c/strong\u003e, 172\u0026ndash;178, DOI: 10.1002/jso.20079 (2004).\u003c/li\u003e\n \u003cli\u003eMervic, L. Time course and pattern of metastasis of cutaneous melanoma differ between men and women. \u003cem\u003ePLoS One\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e7\u003c/strong\u003e, e32955, DOI: 10.1371/journal.pone.0032955 (2012).\u003c/li\u003e\n \u003cli\u003eSchmid-Wendtner, M. \u003cem\u003eet al.\u0026nbsp;\u003c/em\u003eLate metastases of cutaneous melanoma: an analysis of 31 patients. \u003cem\u003eJ Am Acad Dermatol\u0026nbsp;\u003c/em\u003e\u003cstrong\u003e43\u003c/strong\u003e, 605\u0026ndash;609, DOI: 10.1067/mjd.2000.107234 (2000).\u003c/li\u003e\n \u003cli\u003eZhai, X., Kolesnikov, A., Houlsby, N. \u0026amp; Beyer, L. Scaling vision transformers (2022). 2106.04560.\u003c/li\u003e\n \u003cli\u003eGoode A, H. J. J. D. S. M., Gilbert B. Openslide: A vendor-neutral software foundation for digital pathology. \u003cem\u003eJ Pathol Inf.\u0026nbsp;\u003c/em\u003eDOI: doi:10.4103/2153-3539.119005 (2013).\u003c/li\u003e\n \u003cli\u003eXu, H. \u003cem\u003eet al.\u0026nbsp;\u003c/em\u003eA whole-slide foundation model for digital pathology from real-world data. \u003cem\u003eNature\u0026nbsp;\u003c/em\u003e(2024).\u003c/li\u003e\n \u003cli\u003e Ding, J. \u003cem\u003eet al.\u0026nbsp;\u003c/em\u003eLongnet: Scaling transformers to 1,000,000,000 tokens (2023). 2307.02486.\u003c/li\u003e\n \u003cli\u003eGu, Y. et al. Domain-specific language model pretraining for biomedical natural language processing (2020). arXiv: 2007.15779.\u003c/li\u003e\n \u003cli\u003eVaswani, A. \u003cem\u003eet al.\u0026nbsp;\u003c/em\u003eAttention is all you need (2023). 1706.03762.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"WSI, Transformer, BioMedBERT, Melanoma, Prov-Gigapath, Foundation models","lastPublishedDoi":"10.21203/rs.3.rs-6297243/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6297243/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCutaneous melanoma is an aggressive form of skin cancer. Knowledge if a primary melanoma is likely to metastasize is crucial for treatment and survival prediction of melanoma patients. We aimed to develop a predictive tool for determining metastatic potential in primary melanomas utilizing a weakly supervised vision language model. A total of 426 routine stained whole slide images (WSI), along with corresponding histopathological features (Breslow thickness, diameter, presence of dermal mitoses, ulceration and regression), were collected. Of these, 341 samples were used for training and validation, while 85 were reserved as a holdout test set. WSIs were split into patches, and feature embeddings were extracted using Prov-GigaPath. Histopathological features were converted to text, with embeddings generated by BiomedBERT. We developed a multimodal transformer integrating WSIs and histopathological features and conducted an ablation study comparing it to (1) TransMIL using only WSIs and (2) an MLP using only histopathological features. Each model employed a bagging ensemble with five cross-validation models. The multimodal transformer achieved an AUC of 0.887, slightly higher than TransMIL (0.883) and notably better than BertMLP (0.800), highlighting the benefit of including imaging and clinical data for early recognition of melanomas with high metastatic potential.\u003c/p\u003e","manuscriptTitle":"Predicting metastatic potential of primary cutaneous melanomas utilizing weakly supervised vision language model","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-26 09:59:06","doi":"10.21203/rs.3.rs-6297243/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-12-16T16:16:15+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-26T00:18:23+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"111084748093709104749478529091902973046","date":"2025-11-24T20:39:23+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-27T09:00:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"246723846674179294732035599289283190148","date":"2025-09-13T09:58:24+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"155424892022395035770057889835392447035","date":"2025-07-27T07:31:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"188081498428275443338753940796356655758","date":"2025-04-06T20:29:10+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-06T14:30:00+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-06T14:26:43+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-04-04T05:21:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-02T10:02:20+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2025-03-24T16:17:09+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"88d4d6a9-7eef-4087-bb75-89e33f6ff512","owner":[],"postedDate":"March 26th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":46139588,"name":"Health sciences/Diseases/Cancer"},{"id":46139589,"name":"Health sciences/Diseases/Cancer/Skin cancer/Melanoma"},{"id":46139590,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":46139591,"name":"Health sciences/Diseases"},{"id":46139592,"name":"Health sciences/Health care"},{"id":46139593,"name":"Health sciences/Medical research"},{"id":46139594,"name":"Physical sciences/Mathematics and computing/Computer science"}],"tags":[],"updatedAt":"2026-04-07T16:10:32+00:00","versionOfRecord":{"articleIdentity":"rs-6297243","link":"https://doi.org/10.1038/s41598-026-45588-w","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-04-01 15:59:27","publishedOnDateReadable":"April 1st, 2026"},"versionCreatedAt":"2025-03-26 09:59:06","video":"","vorDoi":"10.1038/s41598-026-45588-w","vorDoiUrl":"https://doi.org/10.1038/s41598-026-45588-w","workflowStages":[]},"version":"v1","identity":"rs-6297243","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6297243","identity":"rs-6297243","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}