Progressive Translation of H&E to IHC with Enhanced Structural Fidelity (ProgASP)

Advancing Virtual IHC through Multi-Stage Generative Optimization

Conventional H&E-to-IHC translation models struggle to simultaneously optimize structural fidelity, chromatic accuracy, and cellular boundary clarity.
ProgASP introduces 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 achieved
SSIM of 0.2138 (HER2) and 0.2034 (ER), with FID scores of 49.6 and 40.1 respectively,
demonstrating superior visual quality, biochemical fidelity, and diagnostic interpretability compared to existing methods.

Progressive Translation of H&E to IHC with Enhanced Structural Fidelity

Morphology-Aware Prognostic Model for Five-Year Survival Prediction in Colorectal Cancer (PRISM)

Integrating Morphological Intelligence for Precision Oncology

Traditional computational pathology models often overlook organ-specific morphological cues critical for prognosis.
PRISM introduces a morphology-aware deep learning framework that captures the continuous spectrum of tumor architecture and cellular evolution.
Trained on 8.74 million H&E image patches from 424 stage III colorectal cancer patients, PRISM achieved
0.70 ± 0.04 AUC and a hazard ratio of 3.34 (p < 0.0001) for five-year survival prediction.
It outperformed CRC-specific and large foundation models, demonstrating robust, interpretable, and clinically relevant prognostic power across diverse subgroups.

Morphology-Aware Prognostic Model for Five-Year Survival Prediction in Colorectal Cancer from H&E Whole Slide Images (PRISM)

Streamline Pathology Foundation Model by Cross-Magnification Distillation (XMAG)

Bridging Research-Scale Foundation Models to Real-World Clinical Deployment

Traditional pathology foundation models rely on 20× magnification and process thousands of image patches per slide, making them computationally intensive.
XMAG introduces cross-magnification distillation — transferring diagnostic knowledge from a 20× foundation model (UNI2) to a lightweight 5× student (DINOv2-ViT-B).
This preserves diagnostic fidelity while reducing computational cost by >11×.
Trained on 3.49 million histopathology images from 15 cancer types, XMAG delivers performance within 1% AUC of large models and achieves
real-time inference (8.8 WSIs per minute), establishing a scalable pathway for cost-efficient clinical AI deployment.

Streamline Pathology Foundation Model by Cross-Magnification Distillation (XMAG)

CellEcoNet: Decoding the Cellular Language of Pathology

Teaching AI to Read Tissue Like Language for Lung Adenocarcinoma Recurrence Prediction

Despite surgical resection, up to 70% of invasive lung adenocarcinoma (ILA) patients experience recurrence within five years.
CellEcoNet introduces a spatially aware deep learning framework that interprets histopathology as language —
where cells = words, neighborhoods = phrases, and tissue architecture = sentences.
By modeling these “cellular conversations,” CellEcoNet learns biologically meaningful interactions that drive cancer recurrence.
Fusing cell-level and patch-level embeddings through a novel Cell Patch Fusion module, it achieved
77.8% AUC and a hazard ratio of 9.54, outperforming clinical grading and other computational approaches.
This interpretable framework offers a new lens for precision oncology and recurrence risk stratification.

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

Hyperparameter Optimization and Reproducibility in Deep Learning Model Training

Ensuring Consistency and Transparency in Foundation Model Development for Computational Pathology

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.
The AI4Path Lab 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 PatchCamelyon, LC25000-Lung, and LC25000-Colon benchmarks.

The study revealed that RandomResizedCrop values of 0.7–0.8 yielded superior consistency, and that distributed training without local loss provided the most stable convergence. Learning rates below 5.0e−5 consistently degraded performance, while the LC25000 (Colon) dataset demonstrated the highest reproducibility across runs. These results underscore that reproducible AI in pathology depends not only on open reporting but also on careful experimental design and hyperparameter tuning.

The team offers practical guidelines for achieving reproducibility in large-scale multimodal models, emphasizing transparency in configuration management and control of computational non-determinism. This work serves as a blueprint for building robust, reproducible foundation models in computational pathology.

Hyperparameter Optimization and Reproducibility in Deep Learning Model Training

Predicting Neoadjuvant Chemotherapy Response in Triple-Negative Breast Cancer Using Pre-Treatment Histopathologic Images

Advancing Precision Oncology through Interpretable Pathology-Based Response Prediction

Triple-negative breast cancer (TNBC) presents a major clinical challenge due to its aggressive nature and limited treatment options. This study introduces an attention-based multiple instance learning (MIL) framework that predicts pathologic complete response (pCR) to neoadjuvant chemotherapy directly from pre-treatment H&E-stained biopsy slides. Trained on 174 TNBC patients and externally validated on an independent 30-patient cohort, the model achieved an AUC of 0.86 (internal) and 0.78 (external), demonstrating strong generalization.

Importantly, the model’s attention maps aligned with immune-enriched regions—including PD-L1, CD8⁺ T cells, and CD163⁺ macrophages—achieving mean IoU ≈ 0.46, confirming biological interpretability. This AI-driven pathology framework offers a scalable, interpretable approach for personalized NACT response prediction, potentially reducing overtreatment and guiding precision care in TNBC.

Predicting Neoadjuvant Chemotherapy Response in Triple-Negative Breast Cancer Using Pre-Treatment Histopathologic Images