AACR 2026 Accepted Abstracts

Three AI4Path Lab studies highlight multimodal metastatic-site prediction, rapid HER2 inference from H&E, and real-world clinical deployment in pancreatic cancer.

USCAP 2026 Accepted Abstracts

Five studies from the AI4Path Lab highlight advances in computational pathology — from molecular subtyping and genomic prediction to clinical AI deployment and hematopathology automation.

We’re pleased to announce a new contribution from the AI4Path Lab — Progressive Translation of H&E to IHC with Enhanced Structural Fidelity, led by Yuhang Kang, Ziyu Su, Tianyang Wang, Zaibo Li, Wei Chen, and Muhammad Khalid Khan Niazi.

🔍 What’s the Study About?

Computational stain translation remains a critical challenge in digital pathology. Most existing H&E-to-IHC translation models combine multiple loss terms through simple weighted summation, often producing images that lack optimal structural fidelity, color accuracy, or cellular detail. To address this, the team developed ProgASP, a progressive generative framework that decouples image synthesis into three sequential stages — structure generation, DAB-guided color enhancement, and gradient-refined cell boundary refinement. Evaluated on the MIST dataset for HER2 and ER biomarkers, ProgASP demonstrates superior performance in color consistency, cellular boundary clarity, and overall structural realism.

Figure 1. Overview of ProgASP: A Progressive Generative Framework for Virtual IHC. The model consists of three sequential stages: (1) Structure Generation using ASP loss to ensure robust tissue morphology alignment, (2) DAB-Guided Color Fidelity that leverages 3,3′-DAB channel intensity to enforce biochemical accuracy and protein expression visualization, and (3) Gradient-Guided Cell Boundary Refinement that combines image gradients with DAB-weighted supervision to sharpen cellular boundaries. Each stage receives the output of the preceding module as input, with parameters frozen post-training to preserve feature stability. This hierarchical approach decouples color and structural optimization, enabling targeted refinement of each visual component for clinically interpretable virtual IHC generation.

📊 Key Findings

• Superior structural fidelity: SSIM scores of 0.2138 (HER2) and 0.2034 (ER), outperforming ASP and Stable Diffusion.
• Enhanced color accuracy: DAB-guided loss ensures pixel-level agreement in protein expression intensity and localization.
• Sharper cellular boundaries: Gradient MSE of 0.002234 (HER2) with precise membrane delineation for diagnostic relevance.
• Improved distributional consistency: FID reduced to 49.6 (HER2) and 40.1 (ER), reflecting superior feature preservation.
• Robust progressive training: Sequential stage-wise optimization prevents interference between objectives while maintaining stability.

🔬 These advances highlight that decoupled optimization of structural, chromatic, and morphological features is essential for generating diagnostically reliable virtual IHC images. The progressive framework opens new possibilities for cost-effective, tissue-efficient digital pathology workflows without compromising diagnostic quality.

👥 Meet the Authors

Yuhang Kang, Ziyu Su, Tianyang Wang, Zaibo Li, Wei Chen, and Muhammad Khalid Khan Niazi (PI, AI4Path Lab)

Stay tuned for more studies from AI4Path at the intersection of computational pathology, stain translation, and diagnostic AI.

We’re pleased to announce a new contribution from the AI4Path Lab — Hyperparameter Optimization and Reproducibility in Deep Learning Model Training, led by Usman Afzaal, Ziyu Su, Usama Sajjad, Hao Lu, Mostafa Rezapour, Metin Nafi Gurcan, and Muhammad Khalid Khan Niazi.

🔍 What’s the Study About?

Reproducibility remains a major challenge in foundation model training for histopathology.
Software randomness, hardware non-determinism, and incomplete hyperparameter reporting
often lead to inconsistent results across research groups.

To address this, the team systematically evaluated reproducibility by training a CLIP model on the
QUILT-1M dataset, exploring how different hyperparameter settings and augmentation strategies
influence downstream performance on three key datasets — PatchCamelyon, LC25000-Lung, and LC25000-Colon.

Figure 1. Overview of our joint image–text representation learning framework.
The model jointly trains an image encoder and a text encoder to learn a shared multimodal embedding space by
maximizing the cosine similarity of matched image–text pairs within a batch.
Image patches are processed by the image encoder to obtain latent visual representations
(u₁, u₂, …, u_n), while corresponding textual descriptions are embedded through the text encoder
into feature vectors (v₁, v₂, …, v_n).
Pairwise similarities (u_i·v_j) form a contrastive learning objective that aligns semantically
related histopathology images and diagnostic texts in a unified latent space, enabling the model to capture morphological–linguistic correlations crucial for computational pathology.

📊 Key Findings

• Optimal augmentation: RandomResizedCrop values of 0.7–0.8 outperformed more extreme settings.
• Training stability: Distributed training without local loss produced the most consistent convergence.
• Learning rate sensitivity: Rates below 5.0e−5 consistently degraded model performance.
• Benchmark robustness: The LC25000 (Colon) dataset showed the highest reproducibility across runs.

⚙️ These experiments highlight that achieving reproducible AI in digital pathology depends not only on open reporting
but also on careful experimental design and hyperparameter tuning.
The authors provide practical recommendations for building reliable, reproducible foundation models in the field.

👥 Meet the Authors

Usman Afzaal, Ziyu Su, Usama Sajjad, Hao Lu, Mostafa Rezapour, Metin Nafi Gurcan, and
Muhammad Khalid Khan Niazi (PI, AI4Path Lab)

Stay tuned for more studies from AI4Path at the intersection of foundation models,
computational pathology, and AI reproducibility.

We’re proud to announce a new study from the AI4Path Lab — the release of Morphology-Aware Prognostic Model for Five-Year Survival Prediction in Colorectal Cancer from H&E Whole Slide Images (PRISM), led by Usama Sajjad, Abdul Rehman Akbar, Ziyu Su, Deborah Knight, Wendy L. Frankel, Metin N. Gurcan, Wei Chen, and Muhammad Khalid Khan Niazi.

🔍 What’s the Study About?

Colorectal cancer (CRC) remains the world’s third most common malignancy, with over 150,000 new cases expected in 2025.
While foundation models have advanced computational pathology, their task-agnostic nature often overlooks
organ-specific morphological cues that are vital to understanding tumor biology and predicting patient outcomes.

To bridge this gap, the AI4Path team developed PRISM — an interpretable, morphology-aware prognostic model that
captures the continuous spectrum of phenotypic variability within tumor architecture, reflecting how
malignant evolution unfolds gradually rather than abruptly.

Figure 1. An overview of our PRISM framework.
(a) We first tessellate whole slide images (WSIs) into n non-overlapping patches, each patch undergoing dual feature extraction.
(b) We perform cross-feature interaction between universal pathology features from UNI and morphology-aware features that encode tissue architecture and histopathological patterns.
We then fuse these complementary feature representations f_i,j at the patch level to create comprehensive morphological embeddings.
An attention mechanism computes importance scores for each patch feature based on its prognostic relevance, enabling the model to focus on histologically relevant regions.
We aggregate attention-weighted patch embeddings into a slide-level representation that captures the overall morphological landscape for five-year survival prediction.
(c) During patch feature aggregation, we project features using two neural networks (W_g, W_m) and aggregate the results through W_fusion to obtain morphology-aware patch features.
(d) Based on the predicted probability, we train a time-to-event Cox hazards model or perform risk stratification using the concordance index.

📊 Key Results

• Trained on 8.74 million H&E image patches from 424 stage III CRC patients.
• Achieved superior five-year overall survival prediction (AUC = 0.70 ± 0.04, accuracy = 68.4% ± 4.8%).
• Hazard Ratio = 3.34 (95% CI = 2.28–4.90, p < 0.0001).
• Outperformed existing CRC-specific models by 15% and AI foundation models by ~23%.
• Demonstrated sex-agnostic and treatment-consistent performance across clinical subgroups.

🧠 PRISM highlights the importance of morphology-aware AI — integrating
histological diversity and spatial context to enable more personalized risk assessment
and treatment planning for colorectal cancer patients.

👥 Meet the Authors

Usama Sajjad, Abdul Rehman Akbar, Ziyu Su, Deborah Knight, Wendy L. Frankel, Metin N. Gurcan, Wei Chen, and Muhammad Khalid Khan Niazi (PI, AI4Path Lab)

Stay tuned for more pioneering research from AI4Path, where we continue to push the boundaries
of computational pathology and AI-driven oncology.

We’re excited to share another milestone from the AI4Path Lab — the release of Streamline Pathology Foundation Model by Cross-Magnification Distillation (XMAG), led by Ziyu Su, Abdul Rehman Akbar, Usama Sajjad, Anil V Parwani, and Muhammad Khalid Khan Niazi.

🔍 What’s the Study About?

Foundation models (FMs) have revolutionized computational pathology, but their enormous size and high-magnification image requirements
make them challenging to deploy in real-world clinical workflows.

To overcome these barriers, our team developed XMAG — a lightweight, efficient foundation model built through
cross-magnification distillation. This innovative framework transfers knowledge from a
state-of-the-art 20× magnification teacher to a compact 5× magnification student network,
preserving diagnostic power while dramatically reducing computational cost.

Figure 1. Overview of the XMAG. XMAG transfers knowledge from high-magnification foundation
models to compact low-magnification architectures, achieving comparable diagnostic performance with dramatically improved
processing efficiency across multiple clinical tasks.
(a) Traditional 20× approaches process ~6,000 patches per WSI through large foundation models,
while XMAG operates at 5× magnification with ~500 patches, maintaining diagnostic accuracy with improved efficiency.
(b) Cross-magnification distillation architecture with global/local alignment between teacher (UNI2) and student (DINOv2-ViT-B).
(c) Pretraining dataset: 6,703 WSIs across 15 cancer types (3.49M patches).
(d) XMAG achieves optimal balance between diagnostic performance (AUC) and processing speed (8.8 WSIs/min).
Bubble size indicates parameter count.

💡 Key Innovations

• Cross-magnification knowledge transfer — distills both global and local representations from 20× to 5× magnification.
• Compact architecture — operates entirely at 5×, requiring 11.3× fewer patches per whole slide image (WSI).
• Dual-level distillation — aligns global image features and local spatial token mappings for robust multi-scale learning.
• End-to-end optimization — further boosts the student model to closely match large-scale FM performance.

📊 Results at a Glance

• Trained on 3.49 million histopathology images from public datasets.
• Validated across six clinically relevant tasks spanning multiple cancer types.
• Achieved diagnostic accuracy within 1% of large foundation models.
• Delivered 30× faster processing speed — reaching 8.8 WSIs per minute.
• Confirmed cross-institutional robustness and generalization.

🚀 These results demonstrate that XMAG offers near-FM-level accuracy with real-time performance —
enabling the practical integration of AI into pathology diagnostics even in resource-constrained clinical settings.

🌍 Why It Matters

XMAG redefines what’s possible for foundation models in pathology.
By distilling knowledge across magnifications, it bridges the gap between
research-scale AI and scalable clinical deployment — paving the way for
real-time, cost-efficient pathology AI systems.

👥 Meet the Authors

Ziyu Su, Abdul Rehman Akbar, Usama Sajjad, Anil V Parwani, and
Muhammad Khalid Khan Niazi (PI, AI4Path Lab)

Stay tuned for more pioneering research from AI4Path, where we continue to advance
the boundaries of computational pathology and AI-driven precision medicine.

We’re excited to announce that a groundbreaking new study from the AI4Path Lab, led by
Abdul Rehman Akbar under the supervision of Dr. Muhammad Khalid Khan Niazi,
has been released on arXiv:

Read: CellEcoNet: Decoding the Cellular Language of Pathology with Deep Learning for Invasive Lung Adenocarcinoma Recurrence Prediction

🔍 What’s the Study About?

Despite surgical resection, nearly 70% of invasive lung adenocarcinoma (ILA) patients experience recurrence
within five years. Current clinical grading and staging systems often fail to identify high-risk patients who might
benefit from adjuvant therapy.

To tackle this challenge, our team developed CellEcoNet, a
spatially aware deep learning framework that treats histopathology as a form of language —
where cells are words, cellular neighborhoods form phrases,
and tissue architecture creates sentences.

By modeling this “language of pathology,” CellEcoNet automatically learns context-dependent cellular interactions that
reveal the underlying biology of cancer recurrence.

Figure 1. Schematic overview of the CellEcoNet architecture for invasive lung adenocarcinoma (ILA) recurrence prediction.
The framework integrates tissue-level embeddings (E) and cell-level embeddings (C) through a Cell Patch Fusion module,
producing unified feature representations (f₁ … fₖ). These fused features are aggregated via an attention-based mechanism to generate a global
slide-level embedding (F), which is then used for recurrence classification.
The lower panel illustrates the fusion mechanism, where query, key, and value projections are computed from both patch and cell tokens,
followed by spatially biased attention and a fusion MLP to yield the final fused representation (fₚ).

💡 Key Innovations

• Language-inspired representation of tissue — translating microscopic structure into a contextual model of cellular communication.
• Spatial awareness — integrating both cell-level and microenvironment-level information for precise risk prediction.
• Fair and robust performance — maintaining accuracy across diverse demographic and clinical subgroups.

📊 Results at a Glance

📈 CellEcoNet achieved superior predictive power, significantly outperforming both established
pathological grading systems and state-of-the-art AI methods.

🌍 Why It Matters

CellEcoNet goes beyond prediction — it decodes how the tumor microenvironment “speaks.”
By understanding how subtle variations in cellular arrangements and spatial context encode recurrence risk,
the model opens new avenues for precision oncology, enabling more informed decisions for post-surgical treatment.

👥 Meet the Authors

Abdul Rehman Akbar, Usama Sajjad, Ziyu Su, Wencheng Li, Fei Xing, Jimmy Ruiz, Wei Chen, and
Muhammad Khalid Khan Niazi (PI, AI4Path Lab).

Stay tuned for more pioneering research from AI4Path, where we continue to push the boundaries of computational pathology and AI-driven precision medicine.