Deep Learning Predicts JAK2, Calr TET2, and ASXL1 directly from Whole Slide Marrow Images

Introduction:

Molecular subclassification and risk stratification of Philadelphia chromosome-negative myeloproliferative neoplasms (MPNs) rely on identifying myeloid driver (e.g., JAK2, MPL, CALR) and co-occurrent somatic mutations (e.g., TET2, TP53, ASXL1) [Grinfeld, 2019; Tashkandi, 2024]. This study develops deep learning (DL) models to predict these mutations using Pathomics image analysis of digital hematoxylin and eosin (H&E) whole slide images (WSIs) of MPN bone marrow core biopsies. The translational goal is to use DL Pathomics mutation for efficient, cost-effective screening, allowing rapid identification of abnormal marrows and focusing histologic review on specific regions, thereby enhancing objective diagnostic accuracy and patient care [Kather, 2019]. This result can also identify samples for confirmation with traditional laboratory molecular testing and next-generation sequencing.

Methods:

Following institutional review from Stony Brook University and Weill Cornell Medicine, we identified 271 marrow slides using the WMC MPN research data repository. These slides were from newly diagnosed patients with polycythemia vera (PV; n=83, 31%), essential thrombocytosis (ET; n=83, 31%), myelofibrosis (MF; n=76, 28%), and tumor-negative, normal controls collected for Non-Hodgkin's lymphoma staging (n=35, 13%). The dataset includes H&E WSIs with robust clinical and molecular annotations.

We trained and evaluated six DL algorithms to predict JAK2, CALR, TET2, and ASXL1 by using multiple instance learning (MIL) models [Dolezal, 2024] trained on extracted features (ResNet, HistoSSL, RetCCL, and CTransPath) and end-to-end ResNet and DenseNet models [Liechty, 2022]. A balanced selection strategy allocated 75% of WSIs for training and 25% for testing. WSIs were subdivided into 96K 256 μm square image patches at 10x magnification for computational efficiency and parallel processing. Patch-level results were aggregated into slide-level mutation predictions and used to generate attention heatmaps to examine spatial heterogeneity. Model performance was evaluated by using precision, recall, and F1-score.

Results:

The DL models showed strong performance in predicting mutations from WSIs. For JAK2 mutation prediction, the models achieved precision (P) of 90%, recall (R) of 98%, and F1 of 87%. The corresponding performance for identifying CALR mutations was P=88%, R=98%, and F1=86%. For TET2 mutation prediction, P=78%, R=100%, and F1=78% and for ASXL1 performance was P=76%, R=100%, and F1=76%.

MIL models outperformed DenseNet and ResNet models in terms of precision and F1, and the DenseNet and ResNet models were impacted by smaller training sizes. The RetCCL and CTransPath MIL models excelled, often exceeding average metrics across all prediction tasks. Attention heatmaps revealed that hematopoietic red bone marrow regions correspond to high discriminative power, underscoring the models’ ability to identify key pathologic features.

Conclusions:

Our results demonstrate that Pathomics image analysis can effectively predict myeloid driver and co-occurrent somatic mutations in MPNs. MIL feature extractors that preserve spatial relationships improve model performance. The attention heatmaps generated by our workflow assist pathologists in reviewing and interpreting mutation predictions within the bone marrow microenvironment, facilitating histologic, laboratory, and clinical correlation. This approach can expedite analysis, objectify classification, and select samples for molecular confirmation studies, ultimately improving the identification of abnormalities and focusing attention on specific regions for histologic review.