Primary Central Nervous System Lymphoma (PCNSL) is a particular extra-nodal subtype of large B-cell lymphoma (LBCL) with dismal prognosis. We propose a novel approach using digital pathology to classify PCNSL, assessing the presence of MYD88 L265P mutation and predicting the outcome of PCNSL and LBCL using hematoxylin-eosin (H&E) immunohistochemistry (IHC) samples. We have used different datasets with LBCL and PCNSL for training and validation of our findings with more than 400 patients. The classification with deep learning using H&E to categorize the four molecular clusters in PCNSL showed a median Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.9 [interquartile range, IQR 0.85-0.93]. The spatial transcriptomics and single cell RNAseq analysis showed a high degree of intra-tumoral heterogeneity. In addition, all the partial least squares (PLS) Cox models with cell radiomics or combining cell radiomics with clinical features achieved a C-index higher than 0.9 in all datasets. We were able to correctly classify the four molecular groups obtained with multi-omics approaches only with deep learning applied to H&E samples. These results may facilitate the routine clinical use of this artificial intelligence approach to progressively ease the implementation of precision medicine in LBCL and PCNSL.
Article
Artificial intelligence classifies and predicts the outcome of primary CNS and nodal large B-cell lymphomas
https://doi.org/10.21203/rs.3.rs-3781225/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Primary Central Nervous System Lymphoma (PCNSL) is a non-Hodgkin lymphoma developing in the central nervous system or eyes without systemic spread. While PCNSL is a rare entity compared to other systemic lymphomas, their incidence has been progressively increased during the last two decades and they are associated with dismal prognosis1,2. We and others have participated in the molecular characterization of PCNSL identifying four molecular groups with different somatic mutations, “hot” or “cold” tumor microenvironment (TME) and with strong relationship with the clinical outcome3–5. However, this molecular classification is difficult to implement in clinical practice due to it is necessary to perform several multi-omics analyses to obtain this classification.
One promising solution could be to use artificial intelligence (AI) on the histology samples to obtain down-to-earth and straightforward approaches as a surrogate marker of the molecular groups and to dissect the clinical evolution of aggressive lymphomas. Interestingly, different AI methods have been recently proposed to tailor the prognosis, diagnosis or classification in a broad range of tumors6–9. For example, in B-cell lymphomas these approaches have provided good classification metrics, including large B-cell lymphomas (LBCL)10–12. Furthermore, the analysis of morphological features in segmented cells from LBCL has been recently described as a novel interesting approach to predict survival13. However, these studies have been performed in a limited number of patients, without considering the molecular diversity of LBCL. To the best of our knowledge, PCNSL classification or survival prediction has not been characterized using AI methods.
In the current study, we used different cutting-edge deep and machine learning approaches to obtain the molecular classification and the survival prediction using hematoxylin-eosin (H&E) immunohistochemistry (IHC) in different independent datasets of PCNSL and LBCL. Additionally, we studied the spatial heterogeneity of PCNSL using spatial transcriptomics and were able to highlight the degree of intratumor heterogeneity and correctly identify the molecular groups within PCNSL. We were able to predict the presence of the hotspot myeloid differentiation primary response 88 (MYD88) L265P somatic mutation using AI. Lastly, we assessed the prediction of survival prognosis in PCNSL and LBCL.
Patients cohorts
We used two cohorts of PCNSL (PSL dataset from Paris with 106 patients and Barcelona dataset with 41 patients) diagnosed according to the World Health Organization classification of brain tumors of 2021 in immunocompetent patients, Epstein-Barr negative. The PSL dataset was also molecularly characterized, and we have the four molecular groups for all patients. The PCNSL PSL cohort is composed of 106 patients (respectively 25/20/22/39 patients in clusters 1/2/3/4), Table S1. It has been shown4 that patients in cluster 1 have high proliferation and PRC2 complex activity and are associated with an intermediate outcome prognosis (26 months). Patients in cluster 2 have an immune-cold hypermethylated TME whereas patients in cluster 3 have a heterogeneous one. Patients from these two last clusters have a poor prognosis (respectively 18 and 14 months). Patients of cluster 4 have an immune-hot profile TME. They could be more susceptible to immunotherapy and have the most favorable prognosis (67 months). The PCNSL Barcelona cohort is composed of 41 patients, Table S2. Those with an Eastern Cooperative Oncology Group (ECOG) score greater than two (15 patients) received chemotherapy with or without autologous stem cell transplant (ASCT). The others (with ECOG statuses less or equal to two, 26 patients) received either chemotherapy with or without ASCT. In addition, we used several independent datasets with systemic nodal and extra-nodal LBCL. Patients with LBCL can be divided into two groups: the non-Germinal center subgroup (non-GC) and the Germinal center subgroup (GC). Firstly, the LBCL Rouen cohort, 14 patients are in GC subgroup and 37 in non-GC subgroup, Table S3. Secondly, among the 162 patients from the DLBCL-Morph cohort, 93 were non-GC and 69 GC, Table S4. The tissue microarrays (TMAs) from the DLBCL-Morph dataset were obtained from . Thirdly, in TCGA-DLBCL cohort we analyzed 44 patients, Table S5. We downloaded the diagnostic hematoxylin–eosin (H&E)-stained whole slide through the Genomic Data Commons portal ().
The first PCNSL PSL cohort served as the basis for training of the neural network and machine learning algorithms for the classification of the molecular groups and to assess the survival model. The rest of the cohorts were used to confirm and validate the survival analysis. Details of each cohort are provided in Figure 1, Figure S1 and Table S1-S12.
Each cohort was approved by the local ethics committee and Commission nationale de l'informatique et des libertés (CNIL, for the cohorts in France) in accordance with the Declaration of Helsinki.
Scanning and preprocessing
Slides from the PSL Paris and Rouen cohorts were scanned using a Nanozoomer S60 Hamamatsu (Hamamatsu Photonics) at 40-fold magnification (0.23µm per pixel). TCGA slides were digitized at various institutions participating in the TCGA consortium. In the DLBCL-Morph cohort the stained tissue microarray (TMA) slides were scanned at 40x magnification (0.25µm per pixel) on an Aperio AT2 scanner (Leica Biosystems) in ScanScope Virtual Slide (SVS) format. Slides from the Barcelona cohort were scanned using a Pannoramic Flash III at 40x magnification (0.24µm per pixel).
All images were homogeneously pre-treated using several steps to minimize the artifacts and the density of cells in each image. Smaller image tiles (224 px2) were then generated from the initial scanned images (whole-slide images, WSI). The details of the different pre-processing steps are provided in the supplementary methods. To minimize the potential bias introduced by scanning and staining at different institutions and achieve more robust results, all images were normalized to an external reference image of a different dataset using structure-preserving color normalization as proposed by Reinhard.
Machine and deep learning analyses
We used several AI approaches to analyze the different datasets. For the classification of the four molecular groups in the PSL cohort, we used an EfficientNetv215 neural network architecture using the PyTorch Image Models (TIMM) library (https://github.com/rwightman/timm), pre-trained using ImageNet16. We used a learning rate of 0.00001, with a Label Smoothing Cross Entropy loss as the loss function, AdamW was used as an optimizer, and a momentum of 0.1 was chosen. The default batch size was 10. We used Gradient-weighted Class Activation Mapping (Grad-CAM)17 to visualize the important regions used in the classification.
Additionally, we used constrained-attention multiple-instance learning (CLAM)18 for other classification tasks. The CLAM method was implemented using slidedev (https://slideflow.dev/)19. The model used was clam_sb, with a learning rate of 0.0001, Adam optimizer, 15 epochs and with dropout (p=0.25). Patches were extracted from .ndpi files; tissue was detected with Otsu’s threshold. No other transformation has been applied to the patches. CLAM was used to assess the presence of MYD88 L265P somatic mutation (wild-type or mutated) in PCNSL from PSL and Barcelona cohorts as well as the classification of survival cut-off survival > 12 months. Furthermore, we used EBRAINs dataset20 (https://search.kg.ebrains.eu/instances/Dataset/8fc108ab-e2b4-406f-8999-60269dc1f994) to compare control brain’s slides (without brain tumors) vs PCNSL. Finally, we compared the classification performance to detect the cell of origin (COO) of LBCL, germinal-center (GC) or noGC21. The GC status was obtained using the Hans method22. All deep learning models were performed using Nvidia GPU RTX A5000 with 16Gb.
For the survival analyses, we used another combined approach. Firstly, we used the Hover-Net23 neural network architecture to obtain the nuclear / cell segmentations in each patch for all the patients in the five datasets. Then, we applied PyRadiomics24 to obtain different radiomics features in each nuclear / cell, including like first-order statistics, 2D shape-based, Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Gray Level Dependence Matrix (GLDM) as well as wavelet transformations. Overall, we used 474 radiomic features. Details of the radiomics features are detailed https://pyradiomics.readthedocs.io/en/latest/features.html. We averaged all the radiomics features and the results were scaled using the z-score. Be 
the matrix of the radiomics features (here, p = 474), where 
(n being the number of individuals).
We used a Partial Least Squares (PLS) Cox regression model using the plsRcox package25including all the radiomics features as covariates in this model. This type of approach was used to overcome the potential collinearity between the radiomics features. We used the PSL cohort as a training cohort and the rest of datasets were used to confirm the results. The number of PLS Cox components was obtained with cross-validation using plsRcox R package.
We have included the Minimum information about clinical artificial intelligence modeling (MI-CLAIM) checklist in the Table S13.
Statistical analysis
Performance metrics included recall (sensitivity), true-negative rate (specificity), precision, F1 score, and area under the receiver operating characteristic (AUROC). The mean AUROC either of multiple classes or as a summary of cross-validation for each individual class was calculated using micro- and macro-averaging. Log-rank test was used to compare the survival of two or more groups. The assessment of the prognostic accuracy of PLS Cox models was performed using the concordance index (C-index)26. We used different PLS Cox in each cohort. describe. All statistical tests were two-tailed and p-values <0.05 were considered statistically significant.
AI correctly identifies PCNSL molecular groups, MYD88 status, and COO
The characteristics of the different cohorts of PCNSL and LBCL of patients are described in Figure 1, Figure S1 and Table S1-S12. Briefly, the large majority of PCNSL patients were treated with high-dose methotrexate (HD-MTX) regimens. The LBCL from Rouen cohort exclusively involved nodal LBCL. TCGA-DLBCL was a combination of different nodal and extranodal LBCL and this information is not available in the DLBCL-Morph dataset. All LBCL cohorts were treated using R-CHOP or R-CHOP–like immunochemotherapy.
For our main dataset, PCNSL from PSL, a total of 605 775 tiles was generated from annotations of 106 slides from 106 patients with a mean tile count of 5715 (±76) per patient. For the diagnostic experiments, up to 100 randomly selected tiles per patient were used for training and validation (to keep computing times within a reasonable timeframe).
For our deep-learning models, we used as metrics AUROC curves, which shows the ability of the model to distinguish between classes, the accuracy, to have the number of correct predictions in total to the number of predictions made, and the precision-recall curves summarizing the trade-off between the true positive rate and the positive predictive value with different probability thresholds, especially useful when datasets are imbalanced. We also show confusion matrices to have a global vision of the predictions.
Classification accuracy was 74%. The AUROC for all classes combined was 0.88 (±0.07) using macro-averaging (Figure 2A). For each cluster, the AUROC are respectively [0.879, 0.776, 0.921, 0.958]. Precision-recall curves as indicated in Figure 2A. We represented the attention map of the neural network with Grad-CAM to visualize classification results of all subgroups, Figure 2B. The distribution of samples in training and test is shown in Table S10.
Similarly, we used CLAM neural network method for other classification tasks. We used a subset of patients from PSL and Barcelona cohorts with the status of MYD88 L265P available directly using H&E WSI, Table S10-S11. Remarkably, we obtained good classification results with an AUROC of 0.92 [CI 0.57-0.97], Figure 3A. Then, we used another open-source dataset, EBRAINS with a cohort of brain slides from patients harboring no brain cancer to classify PCNSL vs control brains, Table S12. Unsurprisingly, we obtained virtually perfect classification with AUROC of 0.98 [CI 0.75-0.99], Figure 3B. Lastly, we performed different binary survival classifications using with patients having a survival > 12 months with a moderate classification performance with AUROC of 0.69 [CI 0.42-0.79] Figure 3C. Further details are provided in the Figure S2.
Spatial intra-tumor heterogeneity in PCNSL
Next, we wanted to quantify the intratumor molecular heterogeneity in PCNSL. We used two recently published datasets with spatial transcriptomics using 10X Visium technology and single-cell RNA sequencing (scRNA). We used a list with 100 genes per cluster to identify each cluster4. This gene list has been previously published and it is available in Table S14. We applied the recently described Spatial Transcriptomic Analysis using Reference-Free auxiliarY deep generative modeling and Shared Histology (starfysh)27 algorithm that implements a joint modeling of transcriptomic measurements and histology images, to infer different cell states employing spatial transcriptomics. Interestingly, using tumor enriched PCNSL samples Figure 4A-E, we identified a high degree of intra-tumor heterogeneity, because the 4 molecular groups were present in the same sample, Figure 4F. It is important to highlight that high intratumor heterogeneity and the co-presence of the different groups was found in all the analyzed samples, Figure S3. Additionally, we hypothesized that each molecular cluster of PCNSL may be considered as a particular cell state and we were inspired by the work of Neftel et al28 to represent the correlation of each enrichment score to each cluster in several PCNSL samples analyzed by scRNA, Figure 4G. We could confirm our previous results with bulk RNA-seq data4 showing that cluster 4 is the most strongly expressed throughout the different groups. Conversely, cluster 2 was the least frequent, Figure 4G. Further methodological details are provided in the Supplementary Methods.
Radiomics features predicts the survival in LBCL and PCNSL
Finally, we applied a complementary approach using a neural network to automatically segment all the cells within each patch and then extract a broad range of radiomics features in each cell. The results were scaled and averaged per patient and were used to assess different PLS Cox models. For PSL cohort, we obtained a C-index of 0.73 (std 0.03) by taking only clinical features (KPS, age and gender). By taking only radiomics features, a C-index of 0.97 (std 0.01) was obtained; with all the features (clinical and radiomics), the C-index was 0.98 (std 0.01), Figure 5A. We confirmed our results in an independent PCNSL cohort from Barcelona (we obtained the respectively C-indexes for clinical features only (ECOG, age and gender), radiomics features and all the features of 0.68 (std 0.06), 0.98 (std 0.01) and 0.99 (std 0.01), Figure 5A. We generalized our results in systemic LBCL in three different datasets. For Rouen dataset, we obtained a C-index of 0.60 (std 0.08) for clinical features (GC status and age) and C-indexes of 0.96 (std 0.02) for radiomics features only and all features, Figure 5A. For DLBCL-Morph dataset, we obtained a C-index of 0.64 (std 0.05) for clinical features (GC status, RIPI score and age), 0.97 (std 0.01) for radiomics features and 0.98 (std 0.01) for all features, Figure 5A. Finally, for TCGA dataset, we obtained respectively C-index of 0.55 (std 0.14) (age and gender), 0.96 (std 0.02) and 0.96 (std 0.02), Figure 5A.
The weight of each PLS Cox component was represented according to the 4 clusters of PCNSL, Figure 5B and Figure S4A. All the PLS Cox components were enriched with a broad range of different radiomics features, but there was an overrepresentation of wavelet transformed radiomics within the most relevant ones, Figure 5C and Figure S4B. Interestingly, the 15 PLS Cox components were statistically significant in the PLS Cox multivariate analysis as shown in the forestplot, Figure 6A. Finally, to deepen in the interpretability of these PLS Cox components, we have correlated them with the immune cell deconvolution obtained from bulk RNA-seq. Interestingly, the component 1 was strongly positively correlated with CD4 T cell infiltration and negatively with NK cells and CD8, Figure 6B. Inversely, component 2 was positively correlated with CD8 T cell infiltration, among others, Figure 6B. These results suggest that our cell-based radiomic analysis may capture the different TME network associations and their impact on PCNSL.
The molecular characterization of PCNSL has remarkably improved in the last years with several multi-omics studies showing cold/hot TME and different genetic drivers within the different groups3–5. However, due to the rarity of this entity and the limited amount of material available in the brain stereotactic biopsies other approaches are required to stratify these tumors. Similarly, in systemic nodal or extra-nodal LBCL the role of TME and its impact in the prognosis outcome of patients have been pinpointed in several studies using multi-omics approaches29–31. Equivalently, it is difficult to implement these analyses in clinical practice and new surrogate markers to correctly predict the prognosis of patients are required. Currently, different clinical scales that are usually applied in PCNSL like the Memorial Sloan Kettering Cancer Center (MSKCC) scale32 that combine age and performance status. Accordingly, the International Prognostic Index (IPI) scale33 that includes age, performance status, Ann Arbor stage, serum lactate dehydrogenase (LDH) and the presence of extra-nodal sides is usually applied in LBCL. However, there is room for improvement probably in part due to the LBCL heterogeneity as their survival prediction is still not optimal.
In the current study, we present a novel approach to classify different molecular alterations and predict the prognosis in PCNSL and LBCL. The EfficientNetv2 neural network model obtained a weighted accuracy of up to 74% and a macro-averaged AUROC of up to 90% to predict the molecular clusters. It is important to keep in mind that PCNSL / LBCL exhibit a rather stereotypical histological presentation. However, through the utilization deep learning analysis of H&E IHC, distinct molecular groups have been identified for the first time. Likewise, we were able to classify the activating hotspot mutation MYD88 L265P, one of the most frequent missense mutations in PCNSL3,4 and it is associated with B-cell receptor (BCR) signaling34,35. Interestingly, this pathway could be targeted using Bruton tyrosine kinase inhibitors including different nodal and extra-nodal LBCL36,37. We found promising performance in the classification of this mutation with AUROC of 92%. These results are in line with results in other types of cancers where different AI methods allowed to reasonably well classify a broad range of somatic mutations directly from H&E images. It is tempting to speculate that our study may pave the way to tailor or stratify therapeutic strategies in PCNSL.
Previous studies using a broad range of AI methods have been recently published in different B-cell lymphomas38–44. However, most studies were dedicated to classification tasks (i.e., stratify different types of histological types B-cell lymphomas) and unfortunately, molecular stratification has been rarely considered. Additionally, some efforts have been made to predict survival using H&E slides with less well performing results13. In the present study we used a combined deep learning for segmentation and more traditional machine learning method to the analysis of cell-based radiomics features that were analyzed in different PLS Cox models and we obtained robust and reproducible results in the PSL dataset but also in the different confirmatory PCNSL and nodal LBCL datasets included in this work. Indeed, all the cell-based radiomics PLS Cox models obtained a very good survival prediction with more than 0.9 in C-index metrics, throughout the different datasets with even better when we combined the cell-based nuclear radiomics with clinical features (i.e., age and performance status). However, it is important to bear in mind that throughout the different datasets, the increased performance of the survival models was mainly found in the radiomics alone model with a remarkable improvement in all the datasets and much less pronounced benefit when combining radiomics and clinical features. This could suggest that in nodal and extra-nodal LDBCL cell-based radiomics features may provide enough insight to correctly predict the survival. Noteworthily, in other brain tumors, like gliomas, AI methods also found very good survival prediction performance obtaining a C-index higher than 0.8 in several studies 8,45,46.
Even though this represents the largest study using different types of nodal and extra-nodal LBCL it is important to bear in mind that further prospective validation of our findings is required to be able to generalize the results of the present study. However, it is important to highlight that our findings were confirmed in different independent datasets using different acquisition methods (i.e., different scans met), different methodologies (some studies scans were at 40X and others at 20X) and including heterogeneous datasets from different centers that might be like real life condition.
Finally, this study sets the ground for a personalized and potentially tailored molecular stratification of nodal and extra-nodal LBCL identified with different AI methods. Furthermore, our work stresses the importance of using different AI methods to correctly predict survival in aggressive nodal and extra-nodal LBCL.
Funding
This work was supported by a grant from Investissements d’avenir and by the grant INCa-DGOS-Inserm_12560 of the SiRIC CURAMUS [grant number INCa-DGOS-Inserm_12560], the program ‘investissements d’avenir’ [grant number ANR-10-IAIHU-06], PRT-K/INCa grant LOC-model [2017-1-RT-04], BETPSY project, overseen by the French National Research Agency, as part of the second ‘Investissements d’Avenir’ program [grant number ANR-18-RHUS-0012], Foundation RAM active investments, ARTC foundation (no grant number), Association Amour Amour Amour (no grant number), and ANR JCJC 2023 LOCImm project [ANR-23-CE17-0027-01]. G.M.z.H. was supported by the Deutsche Forschungsgemeinschaft (DFG; grant number ME4050/12-1) and by the Heisenberg program of the DFG (ME4050/13-1). O.M.G. was supported by the Deutsche Forschungsgemeinschaft (DFG; CiM reference number: FF-2016-10).
Disclosure
The authors have declared no conflicts of interest.
Acknowledgements
Histomics platform, the cluster of Paris Brain Institute (ICM), the Onconeurotek brain tumor biobank from ICM. The results published here are in whole or part based upon data generated by the TCGA Research Network: https://www.cancer.gov/tcga . Patients and their families
Data sharing (github, links for the different datasets used in the manuscript)
Code and data examples will be made available under Github https://github.com/Noemie-B/Workflow-Morpho-Histo.git including the training weights of the different classification tasks upon acceptance.
PSL, Rouen and Barcelona’s datasets are not publicly available owing to restrictions in the data-sharing agreement.
ScRNA PCNSL dataset can be found https://zenodo.org/records/7813151
Visium 10X spatial transcriptomics from GEO (https://www.ncbi.nlm.nih.gov/geo/), reference GSE203552.
DLBCL-Morph dataset: https://stanfordmedicine.app.box.com/s/ub8e0wlhsdenyhdsuuzp6zhj0i82xrb1
TCGA-DLBCL dataset: https://portal.gdc.cancer.gov/projects/TCGA-DLBC
EBRAINs dataset: https://search.kg.ebrains.eu/instances/Dataset/8fc108ab-e2b4-406f-8999-60269dc1f994
- Ferreri AJM, Calimeri T, Cwynarski K, Dietrich J, Grommes C, Hoang-Xuan K, et al. Primary central nervous system lymphoma. Nat Rev Dis Primer. June 15, 2023;9(1):1–19.
- Houillier C, Soussain C, Ghesquières H, Soubeyran P, Chinot O, Taillandier L, et al. Management and outcome of primary CNS lymphoma in the modern era: An LOC network study. Neurology. March 10, 2020;94(10):e1027–e1039.
- Radke J, Ishaque N, Koll R, Gu Z, Schumann E, Sieverling L, et al. The genomic and transcriptional landscape of primary central nervous system lymphoma. Nat Commun. May 10, 2022;13(1):2558.
- Hernández-Verdin I, Kirasic E, Wienand K, Mokhtari K, Eimer S, Loiseau H, et al. Molecular and clinical diversity in primary central nervous system lymphoma. Ann Oncol Off J Eur Soc Med Oncol. February 2023;34(2):186–199.
- Alame M, Cornillot E, Cacheux V, Rigau V, Costes-Martineau V, Lacheretz-Szablewski V, et al. The immune contexture of primary central nervous system diffuse large B cell lymphoma associates with patient survival and specific cell signaling. Theranostics. 2021;11(8):3565–3579.
- Pantanowitz L, Quiroga-Garza GM, Bien L, Heled R, Laifenfeld D, Linhart C, et al. An artificial intelligence algorithm for prostate cancer diagnosis in whole slide images of core needle biopsies: a blinded clinical validation and deployment study. Lancet Digit Health. August 1, 2020;2(8):e407–e416.
- Fu Y, Jung AW, Torne RV, Gonzalez S, Vöhringer H, Shmatko A, et al. Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis. Nat Cancer. August 2020;1(8):800–810.
- Chen RJ, Lu MY, Williamson DFK, Chen TY, Lipkova J, Noor Z, et al. Pan-cancer integrative histology-genomic analysis via multimodal deep learning. Cancer Cell. August 8, 2022;40(8):865-878.e6.
- Kather JN, Heij LR, Grabsch HI, Loeffler C, Echle A, Muti HS, et al. Pan-cancer image-based detection of clinically actionable genetic alterations. Nat Cancer. August 2020;1(8):789–799.
- El Hussein S, Chen P, Medeiros LJ, Wistuba II, Jaffray D, Wu J, et al. Artificial intelligence strategy integrating morphologic and architectural biomarkers provides robust diagnostic accuracy for disease progression in chronic lymphocytic leukemia. J Pathol. January 2022;256(1):4–14.
- El Hussein S, Chen P, Medeiros LJ, Hazle JD, Wu J, Khoury JD. Artificial intelligence-assisted mapping of proliferation centers allows the distinction of accelerated phase from large cell transformation in chronic lymphocytic leukemia. Mod Pathol Off J U S Can Acad Pathol Inc. August 2022;35(8):1121–1125.
- Li D, Bledsoe JR, Zeng Y, Liu W, Hu Y, Bi K, et al. A deep learning diagnostic platform for diffuse large B-cell lymphoma with high accuracy across multiple hospitals. Nat Commun. November 26, 2020;11(1):6004.
- Vrabac D, Smit A, Rojansky R, Natkunam Y, Advani RH, Ng AY, et al. DLBCL-Morph: Morphological features computed using deep learning for an annotated digital DLBCL image set. Sci Data. May 20, 2021;8(1):135.
- Board WC of TE. Central Nervous System Tumours [Internet]. [cited October 21, 2023]. Available at: https://publications.iarc.fr/Book-And-Report-Series/Who-Classification-Of-Tumours/Central-Nervous-System-Tumours-2021
- Tan M, Le QV. EfficientNetV2: Smaller Models and Faster Training [Internet]. arXiv 2021 [cited October 21, 2023]. Available at: http://arxiv.org/abs/2104.00298
- Deng J, Dong W, Socher R, Li L-J, Li K, Fei-Fei L. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition 2009 [cited October 21, 2023]; 248–255. Available at: https://ieeexplore.ieee.org/document/5206848
- Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. Int J Comput Vis. February 2020;128(2):336–359.
- Lu MY, Williamson DFK, Chen TY, Chen RJ, Barbieri M, Mahmood F. Data-efficient and weakly supervised computational pathology on whole-slide images. Nat Biomed Eng. June 2021;5(6):555–570.
- Dolezal JM, Kochanny S, Dyer E, Srisuwananukorn A, Sacco M, Howard FM, et al. Slideflow: Deep Learning for Digital Histopathology with Real-Time Whole-Slide Visualization [Internet]. arXiv 2023 [cited October 21, 2023]. Available at: http://arxiv.org/abs/2304.04142
- Roetzer-Pejrimovsky T, Moser A-C, Atli B, Vogel CC, Mercea PA, Prihoda R, et al. The Digital Brain Tumour Atlas, an open histopathology resource. Sci Data. February 15, 2022;9(1):55.
- Alizadeh AA, Eisen MB, Davis RE, Ma C, Lossos IS, Rosenwald A, et al. Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling. Nature. February 3, 2000;403(6769):503–511.
- Hans CP, Weisenburger DD, Greiner TC, Gascoyne RD, Delabie J, Ott G, et al. Confirmation of the molecular classification of diffuse large B-cell lymphoma by immunohistochemistry using a tissue microarray. Blood. January 1, 2004;103(1):275–282.
- Graham S, Vu QD, Raza SEA, Azam A, Tsang YW, Kwak JT, et al. Hover-Net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Med Image Anal. December 1, 2019;58:101563.
- van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res. October 31, 2017;77(21):e104–e107.
- Bastien P, Bertrand F, Meyer N, Maumy-Bertrand M. Deviance residuals-based sparse PLS and sparse kernel PLS regression for censored data. Bioinformatics. February 1, 2015;31(3):397–404.
- Harrell FE, Lee KL, Mark DB. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Stat Med. February 28, 1996;15(4):361–387.
- He S, Jin Y, Nazaret A, Shi L, Chen X, Rampersaud S, et al. Starfysh reveals heterogeneous spatial dynamics in the breast tumor microenvironment [Internet]. bioRxiv 2022 [cited October 21, 2023]; 2022.11.21.517420. Available at: https://www.biorxiv.org/content/10.1101/2022.11.21.517420v1
- Neftel C, Laffy J, Filbin MG, Hara T, Shore ME, Rahme GJ, et al. An Integrative Model of Cellular States, Plasticity, and Genetics for Glioblastoma. Cell. August 8, 2019;178(4):835-849.e21.
- Steen CB, Luca BA, Esfahani MS, Azizi A, Sworder BJ, Nabet BY, et al. The landscape of tumor cell states and ecosystems in diffuse large B cell lymphoma. Cancer Cell. October 11, 2021;39(10):1422-1437.e10.
- Kotlov N, Bagaev A, Revuelta MV, Phillip JM, Cacciapuoti MT, Antysheva Z, et al. Clinical and Biological Subtypes of B-cell Lymphoma Revealed by Microenvironmental Signatures. Cancer Discov. June 2021;11(6):1468–1489.
- Ciavarella S, Vegliante MC, Fabbri M, Summa SD, Melle F, Motta G, et al. Dissection of DLBCL microenvironment provides a gene expression-based predictor of survival applicable to formalin-fixed paraffin-embedded tissue. Ann Oncol. December 1, 2018;29(12):2363–2370.
- Abrey LE, Ben-Porat L, Panageas KS, Yahalom J, Berkey B, Curran W, et al. Primary central nervous system lymphoma: the Memorial Sloan-Kettering Cancer Center prognostic model. J Clin Oncol Off J Am Soc Clin Oncol. December 20, 2006;24(36):5711–5715.
- Ruppert AS, Dixon JG, Salles G, Wall A, Cunningham D, Poeschel V, et al. International prognostic indices in diffuse large B-cell lymphoma: a comparison of IPI, R-IPI, and NCCN-IPI. Blood. June 4, 2020;135(23):2041–2048.
- Ngo VN, Young RM, Schmitz R, Jhavar S, Xiao W, Lim K-H, et al. Oncogenically active MYD88 mutations in human lymphoma. Nature. February 2011;470(7332):115–119.
- Davis RE, Ngo VN, Lenz G, Tolar P, Young RM, Romesser PB, et al. Chronic active B-cell-receptor signalling in diffuse large B-cell lymphoma. Nature. January 2010;463(7277):88–92.
- Hendriks RW, Yuvaraj S, Kil LP. Targeting Bruton’s tyrosine kinase in B cell malignancies. Nat Rev Cancer. April 2014;14(4):219–232.
- Grommes C, Pastore A, Palaskas N, Tang SS, Campos C, Schartz D, et al. Ibrutinib Unmasks Critical Role of Bruton Tyrosine Kinase in Primary CNS Lymphoma. Cancer Discov. August 31, 2017;7(9):1018–1029.
- Syrykh C, Abreu A, Amara N, Siegfried A, Maisongrosse V, Frenois FX, et al. Accurate diagnosis of lymphoma on whole-slide histopathology images using deep learning. Npj Digit Med. May 1, 2020;3(1):1–8.
- Steinbuss G, Kriegsmann M, Zgorzelski C, Brobeil A, Goeppert B, Dietrich S, et al. Deep Learning for the Classification of Non-Hodgkin Lymphoma on Histopathological Images. Cancers. May 17, 2021;13(10):2419.
- Zhang J, Cui W, Guo X, Wang B, Wang Z. Classification of digital pathological images of non-Hodgkin’s lymphoma subtypes based on the fusion of transfer learning and principal component analysis. Med Phys. September 2020;47(9):4241–4253.
- Achi HE, Belousova T, Chen L, Wahed A, Wang I, Hu Z, et al. Automated Diagnosis of Lymphoma with Digital Pathology Images Using Deep Learning. Ann Clin Lab Sci. March 2019;49(2):153–160.
- Orlov NV, Chen WW, Mark Eckley D, Macura TJ, Shamir L, Jaffe ES, et al. Automatic Classification of Lymphoma Images With Transform-Based Global Features. IEEE Trans Inf Technol Biomed. July 2010;14(4):1003–1013.
- Gomolka RS, Korzynska A, Siemion K, Gabor-Siatkowska K, Klonowski W. Automatic method for assessment of proliferation index in digital images of DLBCL tissue section. Biocybern Biomed Eng. January 1, 2019;39(1):30–37.
- Swiderska-Chadaj Z, Hebeda KM, van den Brand M, Litjens G. Artificial intelligence to detect MYC translocation in slides of diffuse large B-cell lymphoma. Virchows Arch. 2021;479(3):617–621.
- MHAttnSurv: Multi-head attention for survival prediction using whole-slide pathology images. Comput Biol Med. May 1, 2023;158:106883.
- Cheerla A, Gevaert O. Deep learning with multimodal representation for pancancer prognosis prediction. Bioinformatics. July 15, 2019;35(14):i446–i454.
There is NO Competing Interest.
- FigSupp1AIPCNSL.pdf
Fig Supp 1
- FigSupp2AIPCNSL.pdf
Fig Supp 2
- FigSupp3AIPCNSL.pdf
Fig Supp 3
- SupplementaryMethodsNatureMedAIPCNSL.docx
Supplementary methods
- excelresumedatasets.xlsx
Supplementary Tables
- Supplementarydata.docx