AI Is Learning to Read Where Immune Cells Sit, Not Just Whether a Tumor Is There
Three preprints apply deep learning to brain, head-and-neck, and lung cancers and converge on one idea: the spatial arrangement of tumor and immune cells carries prognostic signal that single markers miss.
- In a preprint, a deep learning biomarker mapped infiltrative glioma beyond contrast-enhancing margins, reaching 85.5 percent accuracy and an AUC of 0.942 after optimized augmentation. Self-reported m:…7893
- In a preprint, deep learning spatial profiling linked higher MET expression to reduced CD8 T cell infiltration in head and neck cancer, and to response to pembrolizumab plus cabozantinib. Self-reported m:…2465
- In a preprint, a deep learning focal TIL score predicted overall survival in lung adenocarcinoma patients treated with immunotherapy, outperforming the standard PD-L1 biomarker. Self-reported m:…5800
- In a preprint, the lung TIL pipeline reached a concordance correlation of 0.90 for lymphocyte detection and a Dice score of 0.87 for scorable-region segmentation. Self-reported m:…5800
Beyond the visible margin
In a preprint, researchers trained a 3D convolutional network on multiparametric MRI and validated it against biopsy-localized histopathology, producing an iTB map that identifies infiltrative glioma tumor in non-enhancing regions that conventional MRI cannot see. Optimized data augmentation lifted accuracy from 74.5 to 85.5 percent and AUC from 0.859 to 0.942, and the biomarker showed low infiltration probabilities in brain metastases, which served as a biologically grounded non-tumor control. For surgical and radiotherapy planning, the point is concrete: the tumor extends past the edge the scan shows, and a validated map of that extension changes where treatment goes. m:…7893
In a preprint on head and neck squamous cell carcinoma, deep learning spatial profiling found that tumors from Black patients showed reduced CD8 T cell infiltration and increased spatial MET expression compared with tumors from White patients, and that higher MET expression tracked with reduced CD8 infiltration across independent cohorts. High predicted spatial MET expression was also strongly associated with improved response to pembrolizumab plus cabozantinib, particularly in Black patients. The study frames MET not just as a marker of disparity but as a candidate guide to a specific combination therapy. m:…2465
A score that beats the standard
In a preprint on lung adenocarcinoma, a deep learning focal TIL score quantified spatial immune infiltration from routine H&E slides and predicted overall survival in patients treated with immune checkpoint inhibitors, with a hazard ratio of 0.60, outperforming the standard PD-L1 biomarker. The underlying pipeline held up analytically as well, with a concordance correlation of 0.90 for lymphocyte detection and a Dice score of 0.87 for scorable-region segmentation. The practical appeal is that the input is an ordinary pathology slide, not a new assay, so the method could slot into existing workflows if prospective validation confirms it. m:…5800
Taken together, the three preprints describe a shift in what imaging AI is for. The glioma work extends the map, the head-and-neck work links a spatial pattern to a therapy choice and to a disparity in outcomes, and the lung work turns a spatial pattern into a prognostic score that beats the incumbent marker. In each case the advance is not a bigger model but a more careful use of location: where tumor cells are, where immune cells are not, and how the two relate in space. m:…7893 m:…2465 m:…5800
What still needs proving
All three results are preprints, so the sensible stance is interested but conditional. The MET work links spatial MET expression to immune exclusion and to response in the cohorts its authors studied, and the lung TIL work links its focal score to survival in the patients its authors studied. Preprint findings are not peer-reviewed, and each of these signals still needs prospective testing before it changes practice. m:…2465 m:…5800
The second-order question is about equity as much as accuracy. The head-and-neck preprint shows spatial profiling can surface a disparity and point toward a therapy for the group with worse outcomes, but it also shows how easily a model trained on one population could misread another. Any team adopting spatial AI should therefore ask for performance broken down by patient group, not just overall accuracy, and should treat a single-cohort AUC as a starting point rather than a verdict. m:…2465
What is still unresolved
- Whether spatial MET profiling predicts treatment response in a larger prospective trial, which the preprint authors say is still needed.
- Whether the focal TIL score holds up outside the studied lung cancer cohort and across pathology labs with different slide preparation.
- Whether the glioma iTB map changes surgical or radiotherapy outcomes in practice, not just diagnostic accuracy.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.