Deep Learning for Histopathology: WSIs, MIL & Transformers
From Gigapixel Slides to Intelligent Predictions
Skills you will gain:
About Workshop:
Digital histopathology has rapidly evolved with the integration of artificial intelligence, enabling scalable and reproducible analysis of whole-slide images. Traditional pixel-level annotations are costly and time-consuming, making modern approaches like MIL and transformers essential for learning directly from slide-level labels. These methods are now widely used in cancer diagnosis, grading, prognosis, and biomarker discovery.
This workshop provides a structured, dry-lab–only learning experience covering WSI preprocessing, patch extraction, feature learning, MIL pipelines, and transformer architectures tailored for histopathology. Through real datasets and guided hands-on sessions, participants will learn best practices for model evaluation, interpretability, and clinical relevance—bridging the gap between AI research and digital pathology applications.
Aim: This workshop aims to equip participants with practical and theoretical knowledge of applying deep learning techniques to digital histopathology data, with a focus on whole-slide images (WSIs). The program introduces modern approaches such as Multiple Instance Learning (MIL) and Transformer-based models to handle gigapixel pathology images under weak supervision. Participants will gain hands-on experience in building, evaluating, and interpreting AI models for cancer detection and pathology research.
Workshop Objectives:
- Understand WSI formats and challenges in histopathology AI.
- Perform patch extraction and deep feature representation from WSIs.
- Implement MIL models for weakly supervised slide classification.
- Apply transformer architectures for slide-level prediction.
- Evaluate and interpret models using clinically relevant metrics.
What you will learn?
Day 1 – Pathology Foundation Models & Self-Supervised WSI Representation
- Pathology Foundation Models (PFMs) for WSIs (pretrained ViT/CNN for feature extraction).
- Self-supervised & contrastive learning (DINO, MoCo, SimCLR) for robust, stain-invariant features.
- Multi-scale WSI tiling and patch-level representation.
- Building reusable feature stores for downstream tasks.
Hands-On:
- WSI patching + extracting embeddings using a pretrained PFM.
- Visualizing embedding clusters (t-SNE / UMAP).
Day 2 – Transformers & MIL for Slide-Level Tumor Detection
- Attention-based MIL, cross-scale MIL, graph-based MIL for gigapixel WSIs.
- Transformers for WSIs: patch embeddings → token sequences.
- Weak supervision using slide-level labels only.
- Clinical evaluation metrics: AUC, F1, calibration, patient-level CV.
Hands-On :
- Training an attention-based MIL model for tumor classification.
- Building a transformer-based slide classifier.
- Comparing MIL vs transformer performance (ROC, confusion matrix, sensitivity/specificity).
Day 3 – Multimodal, Generative & Explainable AI in Pathology
- Multimodal PFMs combining WSI + clinical/genomic/text embeddings.
- Task-adaptive PFMs: LoRA, adapters, prompt-based fine-tuning.
- Explainable AI: Grad-CAM, attention maps, uncertainty estimation.
Hands-On :
- Building a simple multimodal fusion model (WSI features + clinical variables/text).
- Generating interpretability outputs (Grad-CAM, attention heatmaps).
Intended For :
- Undergraduate/postgraduate degree in Biotechnology, Bioinformatics, Computational Biology, Biomedical Engineering, Pathology, Life Sciences, or related fields.
- Researchers and professionals working in digital pathology, cancer research, medical imaging, or healthcare AI.
- Data scientists and AI/ML engineers interested in medical image analysis.
- Individuals with an interest in the intersection of deep learning and pathology.
Career Supporting Skills
Program Outcomes
- Gain hands-on experience with WSI preprocessing and patch-level learning.
- Build MIL and transformer-based models for slide-level prediction.
- Learn to evaluate models using AUC, sensitivity, and validation strategies.
- Generate interpretable outputs such as attention maps and heatmaps.
- Be prepared to apply AI methods to real-world pathology datasets.