Prof. Maria Vakalopoulou
Title: Computational methods for more accurate and precise digital pathology processing
Abstract: In recent years, the medical research community has devoted significant attention to developing new methods for processing digital pathology slides. In this talk, I will present several novel approaches, benchmarks, and analyses that our group has been working on in this area the last year. I will begin by introducing an efficient and comprehensive benchmark we have developed to evaluate and compare foundation models across a variety of tasks and robustness settings [1]. Next, I will discuss our recent paper proposing a new method for efficient data augmentation in a multi-instance learning framework [2]. Finally, I will outline additional strategies for improving the performance of multi-instance learning models for various clinical endpoints [3].
1. Marza, Pierre, Leo Fillioux, Sofiène Boutaj, Kunal Mahatha, Christian Desrosiers, Pablo Piantanida, Jose Dolz, Stergios Christodoulidis, and Maria Vakalopoulou. "THUNDER: Tile-level Histopathology image UNDERstanding benchmark.NeuRIPS 2025 Benchmark and Dataset track (Spotlight)
2. Boutaj, Sofiène, Marin Scalbert, Pierre Marza, Florent Couzinie-Devy, Maria Vakalopoulou, and Stergios Christodoulidis. "Controllable Latent Space Augmentation for Digital Pathology." ICCV (2025)
3. Lolos, Andreas, Stergios Christodoulidis, Maria Vakalopoulou, Jose Dolz, and Aris Moustakas. "SGPMIL: Sparse Gaussian Process Multiple Instance Learning." arXiv preprint arXiv:2507.08711 (2025).
Prof. Mahdi S. Hosseini
Title:
Foundation Models in Computational Pathology: Clinical Diagnostics and Beyond
Abstract:
The computational advantages of deep learning in AI, integrated with giga-pixel whole slide image (WSI) in pathology, has led to the emergence of a new field called Computational Pathology (CPath) that is poised to transform clinical pathology globally. CPath is dedicated to the creation of automated tools that address and aid steps in the clinical workflow for diagnosing and treatment of cancer diseases. With increasing advancements of foundation models in deep learning, the research focus in this field has expanded to broad range of domains. In this seminar talk I will cover two topics: first, I will present the ongoing research trends in deep learning and computational pathology for developing foundational models. We investigate this from “data preprocessing”, “self-supervised pretraining” and “downstream tasking” for clinical applications. Second, I will delve into the ongoing research that we are currently addressing from Atlas Analytics Lab@Concordia e.g. efficient slide processing algorithms, deep learning optimization, encoder designs, prompt-engineering in vision-language models, and understanding concept alignments in foundation models. I will conclude this talk with the objectives of discussing challenges in clinics and research for future AI developments in CPath and how they can facilitate the transformational changes in clinical pathology.
Bio: Mahdi S. Hosseini is an Assistant Professor in Computer Science and Software Engineering (CSSE) at Concordia University, Adjunct faculty in Pathology at McGill University, and an Affiliate Member of Mila–Quebec AI Institute. He directs the Atlas Analytics Lab (CSSE@Concordia) and holds the Gina Cody Research Chair (GCRC) – Research and Innovation. He is a Lead Area Chair for CVPR 2025-26, Broadening Participation Chair of ICCV 2025, and served as AC for CVPR 2022–24, NeurIPS 2023–25, and ECCV 2024. He received the Amazon Research Award (Generative AI, Fall 2023) for work on auto-populating synoptic histopathology reports. His research at Atlas Analytics advances foundational deep learning and computer vision for efficient computational pathology in industry and clinical settings. He has extensive experience in working with digital pathology industry in different sectors including quality assessment and enhancements of image acquisition systems of Whole Slide Image (WSI) Scanners, development of deep learning solutions in computational pathology by contributing dataset creation and software developments in cancer diagnostics. He currently supervises 1 PDF, 3 PhD, and 4 MSc trainees and has graduated six master’s students since August 2022. His goal, in partnership with hospitals, health-tech, and industry, is to develop Computer-Aided Diagnosis (CAD) systems integrated into clinical pathology for cancer diagnosis, prognosis, and treatment. He has secured over $1.5M in research funding (career development and infrastructure) from NSERC (Discovery), Amazon (ARA), FRQNT-Teams, CFI-JELF, DRAC, IRICoR, and Concordia, supporting team capacity, partner collaboration, and HQP training.
Dr. Vincent Qouc-Huy Trinh
Title: Using spatial biology and computational pathology to bridge clinical pathology and drug discovery in human cancers
Abstract:
This presentation will provide an introduction to clinical pathology and its role in advancing diagnostics, prognostics, and biomarker discovery. We will begin by outlining the workflow of tissue preparation and staining, emphasizing how microscopic examination of Hematoxylin and Eosin (H&E) slides reveals key morphological features of normal and diseased tissues. Attention will be given to histoarchitectural changes, nuclear and cytoplasmic alterations, and stromal interactions that are critical for distinguishing malignant from benign processes. Beyond morphology, we will discuss the tumor microenvironment—including immune infiltration, angiogenesis, and stromal remodeling—and how these biological features influence disease progression and treatment response. A special focus will be placed on biomarker analysis, highlighting emerging approaches such as spatial transcriptomics. By capturing gene expression in the spatial context of tissue architecture, spatial transcriptomics enables integration of molecular data with histopathology, providing unprecedented insights into tumor biology. Finally, we will demonstrate how statistical correlations between molecular profiles and traditional H&E-based pathology can enhance diagnostic accuracy and pave the way for robust biomarker discovery. These approaches underscore the growing convergence of pathology, molecular biology, and computational methods in modern precision medicine
Bio: Dr. Vincent Quoc-Huy Trinh is an assistant professor at the University of Montreal, staff pathologist at the University of Montreal Health Center, and researcher at the Institute for Research in Immunology and Cancer in Montreal where he also serves as director of the Pathology core. He obtained his medical degree from the University of Montreal and trained as an anatomical pathology resident at the University of Montreal from 2014 to 2019, during which he also completed a Master's in genetics. He was then a GI and liver clinical fellow from 2019-2020 at Vanderbilt University and research instructor from 2020 until 2022 at the same institution with Youngmin Lee, Kathleen DelGiorno, Marcus Tan, and Robert Coffey. His research team is focused on the transition of digestive lesions from precursors to invasive cancer using pathology, spatial biology, deep-learning, and mouse models.
Prof. Nicolas Thome
Title : Uncertainty Quantification and adaptation of Vision-Language Models (VLMs)
Abstract: Recently, Vision-Language Models (VLMs) pre-trained on large scale datasets enable the design of universal models, especially by their capacity to perform zero-shot classification. However, reliably quantifying the uncertainty of VLMs remains an open challenge. In this talk, I present two recent contributions to make progress towards this direction.
Firstly, I introduce ViLU, a new Vision-Language Uncertainty Quantification framework that is tailored for post-hoc failure prediction of VLMs. VilU accurately models the uncertainty coming from both images and texts. It uses a binary classifier to distinguish correct from incorrect predictions, that provides a consistent generalization of the Maximum Concept Matching (MCM) score. Experiments conducted on classification and large-scale image-caption datasets show the relevance of the proposed approach, which opens the way to zero-shot UQ with VLMs.
The second contribution relates to Test-Time-Adaptation (TTA). I introduce CLIPTTA, a gradient-based TTA method for VLMs that leverages a soft contrastive loss aligned with CLIP’s pre-training objective. We provide a theoretical analysis of CLIPTTA ’s gradients, showing how its batch-aware design mitigates the risk of pseudo-label drift and class collapse - two two common failure modes of entropy minimization methods with VLMs. We further extend CLIPTTA to the open-set setting, where both in-distribution (ID) and out-of-distribution (OOD) samples are encountered. Extensive experiments on 75 datasets spanning diverse distribution shifts show the very good performances of CLIPTTA and its stable performance across diverse shifts.
Bio:
Nicolas Thome is a professor at Sorbonne University within the robotics lab (ISIR, Institut des Systèmes Intelligents et de Robotique), and a senior member of the Institut Universitaire de France (IUF). His research focuses on “Machine Learning for Robotics” (MLR), and he leads the MLR project-team at ISIR. His current work addresses multimodal learning with generative AI, physics-informed hybrid control, and the robustness of ML models for human–robot interactions, with applications in surgical robotics.
Dr. Ahmed Alagha
Title:
A Scalable AI-powered Tool for Efficient Whole Slide Image Analysis
Abstract:
Whole-slide images (WSIs) are central to computational pathology, yet their gigapixel scale and large non-informative regions require efficient preprocessing to enable downstream analysis. Tissue segmentation is a critical step, but existing workflows often rely on patch-first strategies that exhaustively tile slides and later discard background patches, leading to significant computational overhead. Other widely adopted methods depend on heuristic filters or legacy deep learning architectures such as U-Net, which struggle to generalize across diverse tissues, stains, and scanners. Moreover, existing methods rarely provide optimized implementations for patchification, often performing it sequentially in ways that are time-consuming, inefficient, and lacking parallelization support. In this work, we introduce a plug-and-play tool that integrates segmentation and patchification into a single efficient workflow. We curate a large dataset of annotated WSI thumbnails and fine-tune an efficient segmentation model using parameter-efficient finetuning of SAM2 to detect tissue regions at the thumbnail level, with masks extrapolated to full-resolution WSIs for precise patch extraction without exhaustive patchification. The tool is to be released with open access, providing a scalable, generalizable, and easy-to-use resource that reduces preprocessing cost while enabling robust downstream tasks such as multiple instance learning and representation learning.
Bio: Ahmed Alagha is currently a Postdoctoral Fellow in the Department of Computer Science and Software Engineering at Concordia University, Montreal, QC, Canada. He received his Ph.D. in Information Systems Engineering from Concordia University, QC, Canada, and the B.Sc. and M.Sc. degrees in Electrical and Computer Engineering from Khalifa University, Abu Dhabi, UAE, where he also worked as a Research Associate. He is a recipient of the prestigious FRQNT doctoral and postdoctoral awards. His research interests include computational pathology, vision-language models, multiagent systems, deep reinforcement learning, and sensing technologies. He is a reviewer of server prestigious conferences and journals.
Yousef Kotp
Title:
Slide HuBERT: Neighborhood-Aware Self-supervised Pretraining in Computational Pathology
Abstract:
Whole-slide images are gigapixel, yet most representation learning in computational pathology remains tile-centric where patches are embedded independently with ever-heavier backbones and only later pooled by MIL, implicitly assuming that a tile’s meaning is intrinsic. In practice, semantics are context-conditioned, a gland abutting tumor nests suggests invasion where the same gland beside adipose may indicate a clean margin; in isolation it is ambiguous, so pouring all compute into single-tile encoders both misaligns with pathology and scales poorly. Slide HuBERT reframes pretraining around regional context where instead of over-investing the compute per tile, we distribute compute across a local window of neighboring patches. A lightweight patch encoder extracts features once, a context transformer then refines them jointly within spatial grids and performs HuBERT-style masked cluster prediction using multi-granularity pseudo-labels, encouraging features that reflect how tile representations change with their neighborhood. This design amortizes cost over regions, captures slide-aware cues unavailable to isolated tiles, and produces compute-efficient embeddings that transfer cleanly to MIL and other downstream tasks by operationalizing the core insight that in pathology, context defines content.
Bio:
Yousef Kotp is a master’s student in Computer Science at Concordia University under the supervision of Dr. Mahdi S. Hosseini. He obtained his bachelor’s degree in Computer and Communication Engineering from Alexandria University in Egypt. He is currently conducting research on foundation modeling for computational pathology in Atlas Analytics Lab at Concordia.
Balamurali Murugesan.
Title: Constrained-Based Calibration for Reliable Deep Learning Models
Abstract: Deep neural networks are increasingly deployed in high-stakes decision-making systems, where reliable confidence estimation is just as important as predictive accuracy. However, calibration—how well a model’s predicted probabilities reflect true likelihoods—remains a persistent challenge, especially in complex tasks like semantic segmentation and in models adapted from foundation models such as CLIP.
In this talk, I will present two complementary lines of work addressing miscalibration in these settings. First, we revisit calibration in deep segmentation networks, where traditional methods largely ignore spatial structure and treat pixels independently. We introduce Neighbor Aware CaLibration (NACL), a principled approach that imposes equality constraints on logit values across neighboring pixels. Unlike prior methods like Spatially Varying Label Smoothing (SVLS), NACL offers explicit control over the strength and influence of these constraints. Our experiments show that NACL significantly improves calibration while preserving segmentation quality, and is broadly applicable across model architectures.
Second, we address a novel and under-explored problem: miscalibration in CLIP-based model adaptation, particularly under distributional shifts. We show that popular adaptation methods—Adapters, Prompt Learning, and Test-Time Adaptation—can severely degrade the zero-shot model’s calibration due to expanded logit ranges. To counter this, we propose a simple, model-agnostic fix: rescaling logits based on zero-shot predictions, with three variants that can be applied during or after adaptation. Our method consistently improves calibration across OOD benchmarks while maintaining discriminative performance. Together, these works highlight the importance of structure-aware and adaptation-aware calibration strategies for trustworthy AI.
Speaker Bio: Balamurali Murugesan is a Doctoral student at École de technologie supérieure (ÉTS), Montreal, under the supervision of Dr. Jose Dolz and Dr. Ismail Ben Ayed. His current research focuses on building reliable and calibrated medical image segmentation models and vision-language adapters, Prior to his PhD, he was a research scholar at the Indian Institute of Technology, Madras (IIT-M), where he worked with Mr. Keerthi Ram and Dr. Mohanasankar Sivaprakasam on improving image reconstruction quality in medical imaging systems.
He has published in several prestigious conferences including ECCV, WACV, MICCAI, ISBI, MIDL, SPIE, EMBC, and MeMeA, and in high-impact journals such as MedIA, TMLR, CMIG, NeuroComp, and BSPC. He was a finalist in the Thesis Madness event at MICCAI 2024 and received the Institute Research Award for his master’s thesis work at IIT-M. He has also served as reviewer for several key venues including TMI, NeurIPS, ICML, MICCAI, MIDL, ML4H, CHIL, AIME, CIBM.
Professionally, he has completed full-time research internships at Microsoft (2023) and Amazon (2024, 2025), contributing to projects involving transcription accuracy using custom vocabularies, context compression for large language models, and generating editable diagrams from multimodal inputs. He has also served as a consultant for several startups, assisting in solving key product and research challenges.
Earlier in his career, he worked as a software developer at HTIC, where he built full-stack tools for brain image annotation and visualization, and as a research assistant at IIT-M, focusing on device development for action and character recognition.
Ghassen Baklouti
Title: Test-Time Adaptation of Medical Vision-Language Models
Abstract: Integrating image and text data through multi-modal learning has gained significant attention in medical imaging research, building on its success in computer vision. While substantial progress has been made in developing medical foundation models and enabling their zero-shot transfer to downstream tasks, the potential of test-time adaptation remains largely underexplored in this domain. Inspired by recent advances in transductive learning and parameter-efficient fine-tuning, we investigate test-time adaptation of medical vision-language models. Our method leverages information-theoretic principles, maximizing the mutual information between the visual inputs and the text-based class representations, while minimizing a Kullback-Leibler divergence term penalizing deviation of the predictions from the zero-shot outputs. Building on this foundation, we introduce the first structured benchmark for test-time adaptation of medical vision-language models, exploring strategies tailored to the unique challenges of medical imaging. Our extensive experiments include two medical modalities, three specialized foundation models, six downstream tasks, and multiple state-of-the-art test-time adaptation methods, demonstrating significant performance improvements and establishing a new benchmark for this emerging field. The code is available at: https://github.com/FereshteShakeri/TTAMedVLMs.