Abstract B048: Large Scale, AI-Enabled, Spatial Signal Processing of Breast Cancer Pathology Identifies Consensus Tissue Structures Related to Biology and Outcomes
Bibliographic record
Abstract
Abstract Breast cancer remains a leading cause of cancer-related mortality worldwide, underscoring the critical need for innovative diagnostic and prognostic approaches. Spatial evaluation of tissue structure in breast cancer provides valuable insights into the tumor microenvironment, including cellular organization, stromal interactions, and molecular heterogeneity. However, large-scale, high-resolution profiling of tumor structure is not financially nor logistically feasible. In this study, we sought to interrogate spatial signal organization in a large set of breast pathology images to identify common structural features and associated biology, in an unsupervised manner, and connect the presence of these features to clinically relevant endpoints. To accomplish this objective, we utilized a large cohort (n=1988) of normal breast and breast cancer biopsy digital, whole-slide, pathology images (WSI), from TCGA (n=1058), OSU’s TCC (n=401), and GTEx (n=529). WSI processing and QC included color normalization and artifact segmentation. Quality tissue areas were tiled into non-overlapping 224mm x 224mm boxes of 1mm/px and feature embeddings were subsequently extracted from tiles using a deep learning pathology foundation model (CTransPath). Each embedding was subsequently converted to spatial Fourier coefficients (FCs), specific to each sample, using spatial graph Fourier transform, and converted to feature spatial-coordination maps generated from cosine similarity of embedding FCs. To identify conserved tissue structures, we developed a customized graph-neural-network (GNN), trained to identify a common latent feature space of all tile embeddings, based on spatial similarity among all samples, and performed Louvain clustering of the resultant feature latent space. We identified 33 conserved tissue structures (FTU) consisting of 15 to 78 spatially coordinated image features each. We quantified the feature prominence in each sample through feature spatial aggregation and compared FTU quantitation to gene expression programs and clinical features in 1459 breast cancers from TCGA and TCC. Almost half of the FTU’s correlated strongly with molecular subtypes (ER, PR, HER2 status, p<0.05) and every identified FTU associated with a gene expression module. This included gene expression modules associated with lymphocyte infiltration, which paired immune related gene expression modules with at least two FTUs. Additionally, one FTU (FTU-14) correlated significantly (p=0.03) with microbial abundance, derived from RNA-Seq, suggesting bacteria may impact local tissue structure, detectable in H&E images. This work highlights the growing appreciation for spatial ecology in tumors and particularly breast cancer tissue structure. We also demonstrates that computationally-identifiable, conserved tissue structures in breast cancer, derived from digital pathology, can serve as a biomarker for diagnosis and prognosis, with direct relationship to biology and clinical course. Citation Format: Jordan E. Krull, Mirage Modi, Yi Jiang, Karthik Chakravarthy, Daniel Spakowicz, Qin Ma. Large Scale, AI-Enabled, Spatial Signal Processing of Breast Cancer Pathology Identifies Consensus Tissue Structures Related to Biology and Outcomes [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr B048.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.000 |
| Open science | 0.000 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".