Nikos Tsiknakis

Treatment decisions in early breast cancer depend on accurate prognostication and on therapy-relevant biomarkers, both of which rely largely on examination of routine hematoxylin and eosin (H&E)-stained tissue. Manual assessment, however, is subject to observer variability and cannot exhaustively quantify tissue features or detect complex morphological associations. Advances in artificial intelligence have therefore prompted the development of computational tools that extract additional information from routine sections. This thesis investigates whether deep learning applied to H&E whole-slide images (WSIs) can improve prognostication beyond established clinicopathological factors and enable inference of therapy-relevant biomarkers in breast cancer. Four studies are presented, addressing (i) the design of a general modeling framework, (ii) the interchangeability of models for assessing an immune biomarker (i.e., tumor-infiltrating lymphocytes), (iii) the inference of molecular alteration from morphology, and (iv) the direct modeling of patient survival.In the first study, we developed a model-agnostic, multiresolution multiple-instance learning (MIL) framework for WSI analysis, that aggregates features across magnifications independently of the underlying encoder or classifier. Nine MIL models were evaluated with both a domain-specific and an ImageNet-pretrained feature encoder on grade, TP53 mutation status, and survival prediction. Aggregating high- and low-magnification features by concatenation or a learnable linear transformation consistently outperformed single-resolution baselines across models and tasks, indicating that multiresolution representation is broadly beneficial rather than tied to a particular architecture.The second study examined whether independently developed models for automated tumor-infiltrating lymphocyte (TIL) scoring can be used interchangeably in triple-negative breast cancer (TNBC). Ten machine-learning and deep-learning approaches were compared against pathologist-read stromal TILs for analytical and prognostic validity. Agreement with pathologist scoring varied across models and declined from internal to external validation. Despite this variability, eight of ten models retained significant and overlapping prognostic associations. The findings indicate that model choice affects the scores themselves, cautioning against treating such models as interchangeable in the absence of a common benchmark.In the third study, we developed a model to infer TP53 mutation status directly from H&E WSIs, with the aim of providing a histology-based surrogate marker with prognostic value. On a held-out test set, the model discriminated mutated from wild-type tumors with an AUC of 0.83 (95% CI 0.79-0.88), while in an independent, endocrine-treated ER+ cohort, the image-derived score was associated with overall survival (HRadj = 3.88, 95% CI 1.37-10.99; p=0.011). The results provide proof-of-concept for a histology-based TP53 marker that carries independent prognostic information.Building on the third study, the fourth study modeled overall survival directly from H&E WSIs, rather than through a molecular proxy. After adjustment for established prognostic factors, the resulting risk score stratified patients into high- and low- risk groups in the internal test set (adjusted HR, 2.17; 95% CI, 1.09-4.33; p = 0.028) and in two external cohorts: Stockholm (adjusted HR, 1.79; 95% CI, 1.23-2.62; p = 0.003) and PANTHER (adjusted HR, 1.68; 95% CI, 1.22-2.32; p = 0.001). The score also retained prognostic value in the ER+/HER2-subgroup and for recurrence endpoints. These findings suggest that a survival-trained H&E biomarker can provide risk stratification beyond established clinicopathological factors.Taken together, these studies show that deep learning applied to routine H&E WSIs can recover prognostic and molecular information relevant to breast cancer, and they address this question at several levels, from framework-level design choices to task-specific biomarkers. They also identify methodological factors that condition the reliability of such models. The clinical role of these approaches, including any bearing on therapy selection, remains to be established through head-to-head comparison with existing tools and prospective validation.List of scientific papersI. Tsiknakis N, Manikis G, Tzoras E, Salgkamis D, Vidal JM, Wang K, Zaridis D, Sifakis E, Zerdes I, Bergh J, Hartman J, Acs B, Marias K, Foukakis T. Unveiling the power of model-agnostic multiscale analysis for enhancing artificial intelligence models in breast cancer histopathology images. IEEE Journal of Biomedical and Health Informatics, 2024 , 28(9), pp.5312-5322. PMID: 38865229. https://doi.org/10.1109/JBHI.2024.3413533 II. Vidal JM*, Tsiknakis N*, Staaf J, Bosch A, Ehinger A, Nimeus E, Salgado R, Bai Y, Rimm DL, Hartman J, Acs B. The analytical and clinical validity of AI algorithms to score TILs in TNBC: can we use different machine learning models interchangeably ?. EClinicalMedicine. 2024 Dec 1;78. PMID: 39634035. https://doi.org/10.1016/j.eclinm.2024.102928III. Tsiknakis N, Salgkamis D, Tzoras E, Wang K, Liu X, Manikis G, Sifakis E, Marias K, Acs B, Schwab M, Büttner F, Schroth W, Bergh J, Matikas A, Zerdes I, Foukakis T. A prognostic deep learning biomarker based on TP53 mutation status in breast cancer histology. [Manuscript]IV. Tsiknakis N, Salgkamis D, Tzoras E, Manikis G, Liu X, Marias K, Acs B, Hartman J, Hellström M, Johansson H, Andersson A, Loibl S, Untch M, Denkert C, Jank P, Zerdes I, Matikas A, Bergh J, Foukakis T. Learning survival from morphology: development and external validation of a deep learning H&E biomarker in early breast cancer. [Manuscript]* Equal contribution.

Background: Prostate cancer (PCa) diagnosis using MRI is often challenged by lesion variability. Methods: This study introduces Simplatab, an open-source automated machine learning (AutoML) framework designed for, but not limited to, automating the entire machine Learning pipeline to facilitate the detection of clinically significant prostate cancer (csPCa) using radiomics features. Unlike existing AutoML tools such as Auto-WEKA, Auto-Sklearn, ML-Plan, ATM, Google AutoML, and TPOT, Simplatab offers a comprehensive, user-friendly framework that integrates data bias detection, feature selection, model training with hyperparameter optimization, explainable AI (XAI) analysis, and post-training model vulnerabilities detection. Simplatab requires no coding expertise, provides detailed performance reports, and includes robust data bias detection, making it particularly suitable for clinical applications. Results: Evaluated on a large pan-European cohort of 4816 patients from 12 clinical centers, Simplatab supports multiple machine learning algorithms. The most notable features that differentiate Simplatab include ease of use, a user interface accessible to those with no coding experience, comprehensive reporting, XAI integration, and thorough bias assessment, all provided in a human-understandable format. Conclusions: Our findings indicate that Simplatab can significantly enhance the usability, accountability, and explainability of machine learning in clinical settings, thereby increasing trust and accessibility for AI non-experts.

Developing AI models for digital pathology has traditionally relied on single-scale analysis of histopathology slides. However, a whole slide image is a rich digital representation of the tissue, captured at various magnification levels. Limiting our analysis to a single scale overlooks critical information, spanning from intricate high-resolution cellular details to broad low-resolution tissue structures. In this study, we propose a model-agnostic multiresolution feature aggregation framework tailored for the analysis of histopathology slides in the context of breast cancer, on a multicohort dataset of 2038 patient samples. We have adapted 9 state-of-the-art multiple instance learning models on our multi-scale methodology and evaluated their performance on grade prediction, TP53 mutation status prediction and survival prediction. The results prove the dominance of the multiresolution methodology, and specifically, concatenating or linearly transforming via a learnable layer the feature vectors of image patches from a high (20x) and low (10x) magnification factors achieve improved performance for all prediction tasks across domain-specific and imagenet-based features. On the contrary, the performance of uniresolution baseline models was not consistent across domain-specific and imagenet-based features. Moreover, we shed light on the inherent inconsistencies observed in models trained on whole-tissue-sections when validated against biopsy-based datasets. Despite these challenges, our findings underscore the superiority of multiresolution analysis over uniresolution methods. Finally, cross-scale analysis also benefits the explainability aspects of attention-based architectures, since one can extract attention maps at the tissue- and cell-levels, improving the interpretation of the model's decision. The code and results of this study can be found at github.com/tsikup/multiresolution_histopathology .

The tumor immune microenvironment (TIME) is an important player in breast cancer pathophysiology. Surrogates for antitumor immune response have been explored as predictive biomarkers to immunotherapy, though with several limitations. Immunohistochemistry for programmed death ligand 1 suffers from analytical problems, immune signatures are devoid of spatial information and histopathological evaluation of tumor infiltrating lymphocytes exhibits interobserver variability. Towards improved understanding of the complex interactions in TIME, several emerging multiplex in situ methods are being developed and gaining much attention for protein detection. They enable the simultaneous evaluation of multiple targets in situ, detection of cell densities/subpopulations as well as estimations of functional states of immune infiltrate. Furthermore, they can characterize spatial organization of TIME—by cell-to-cell interaction analyses and the evaluation of distribution within different regions of interest and tissue compartments—while digital imaging and image analysis software allow for reproducibility of the various assays. In this review, we aim to provide an overview of the different multiplex in situ methods used in cancer research with special focus on breast cancer TIME at the neoadjuvant, adjuvant and metastatic setting. Spatial heterogeneity of TIME and importance of longitudinal evaluation of TIME changes under the pressure of therapy and metastatic progression are also addressed.

Pollen identification is an important task for the botanical certification of honey. It is performed via thorough microscopic examination of the pollen present in honey; a process called melissopalynology. However, manual examination of the images is hard, time-consuming and subject to inter- and intra-observer variability. In this study, we investigated the applicability of deep learning models for the classification of pollen-grain images into 20 pollen types, based on the Cretan Pollen Dataset. In particular, we applied transfer and ensemble learning methods to achieve an accuracy of 97.5%, a sensitivity of 96.9%, a precision of 97%, an F1 score of 96.89% and an AUC of 0.9995. However, in a preliminary case study, when we applied the best-performing model on honey-based pollen-grain images, we found that it performed poorly; only 0.02 better than random guessing (i.e., an AUC of 0.52). This indicates that the model should be further fine-tuned on honey-based pollen-grain images to increase its effectiveness on such data.