A new artificial intelligence (AI) model is taking a multitask approach to computational pathology, using routine histopathology slides to generate multiple types of information about cancer. The model was designed to produce seven outputs from a single whole-slide image, including tumor type, TP53 mutation status, TP53 RNA expression, and survival-related outcomes.
The study, published in The American Journal of Pathology, evaluated the model across 32 types of solid cancer, highlighting an approach that could potentially extract diagnostic, molecular, and prognostic information within the same analytical framework.
One slide, multiple outputs
The Vision Transformer model was trained using more than 11,000 primary tumor cases from the Pan-Cancer Atlas. The dataset included corresponding somatic mutation, RNA-sequencing, and clinical outcome data.
In an independent validation set of 1,729 slides, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.766 for detecting TP53 mutations across the 32 tumor types. The researchers also found that the model could infer TP53 RNA expression and tumor taxonomy directly from whole-slide images.
TP53 was a particular focus because it is one of the most frequently altered tumor suppressor genes across human cancers. Mutations in the gene can influence tumor growth and treatment resistance, making molecular characterization relevant to cancer diagnosis and treatment.
“Currently, most deep learning-based models are used for single-model concepts; one model for one task. We developed a single model that can generate seven outputs simultaneously from the whole histopathology image, including TP53 mutation status, TP53 RNA expression, tumor type, and survival-related outcomes,” said co-lead investigator Alex W. Hewitt, PhD, of the Menzies Institute for Medical Research and School of Medicine, University of Tasmania, in a press release.
Learning without exhaustive annotation
The researchers used weakly supervised learning rather than requiring exhaustive manual annotation of tumor regions. Molecular labels such as TP53 mutation status were available at the slide level, allowing the model to learn from image patches without researchers having to identify every region containing relevant morphological features.
Input patches were extracted at 6× downsampling, corresponding to an approximate magnification of 6.7× relative to the original WSI resolution (approximately 40×). Slide-level attention across four WSIs was randomly taken from the independent set. Each row corresponds to one slide, showing the thumbnail, overlay attention, and the highest- and lowest-attention patches at 40× magnification, with boxes covering a large area of the tissue (the highest- and lowest-attention patches are at the center of the boxes’ tissues) to highlight the selected area of the WSIs so that the boxes are visible. The model predicted cancer type, TP53 mutation status, TP53 RNA expression levels, and clinical outcomes for overall survival (OS) and progression-free interval (PFI) events, measured in months. A: Cancer: rectum adenocarcinoma | TP53: 0 [expression (expr) 10.07] | OS: 1 (65.5 months) | PFI: 1 (44.8 months). B: Cancer: head and neck squamous cell carcinoma | TP53: 0 (expr 10.86) | OS: 0 (49.2 months) | PFI: 0 (29.7 months). C: Cancer: brain lower-grade glioma | TP53: 1 (expr 10.31) | OS: 0 (38.2 months) | PFI: 0 (33.0 months). D: Cancer: prostate adenocarcinoma | TP53: 0 (expr 10.03) | OS: 1 (48.5 months) | PFI: 1 (41.1 months).
The American Journal of Pathology / Chaurasia et al., CC BY-NC-ND
According to the researchers, this approach could help address the difficulty and cost of creating detailed annotations for large whole-slide images.
The model is not intended to replace molecular testing. Instead, the researchers describe it as a potential screening, prioritization, or decision-support tool that could help identify patients who may benefit from confirmatory testing or support diagnostic triage where genomic testing is limited.
The findings point toward a broader role for computational pathology in which the information contained within a routine tissue image could extend beyond morphology to include molecular and prognostic signals. Further development and validation would be needed before such an approach could be incorporated into routine clinical workflows.
“Accurate molecular profiling from routine histopathology slides, already widely used in cancer care, could transform clinical oncology,” said Hewitt. “This new AI-based model integrates diagnostic, molecular, and prognostic tasks, and could help clinicians obtain more information from existing pathology workflows, ultimately supporting more accessible precision cancer care and early intervention.”
This article was created with the assistance of Generative AI and has undergone editorial review before publishing.
