Weekly Research Digest

Oncology AI Weekly

2026-05-21  ·  5 papers reviewed

Editor's Note

This week's standout story is the external validation of a mammography-derived AI model achieving a 10-year AUC of 0.72 across diverse US and Swedish populations — a genuine step toward AI-guided breast cancer primary prevention. It arrives alongside a sobering scoping review revealing that 371 xAI studies in cancer imaging remain largely unvalidated and undeployed, framing a critical tension: we can build impressive models, but the path from algorithm to clinic is still poorly paved. As prognostic ML tools proliferate across rare and common malignancies alike, standardisation, reproducibility, and prospective validation must become non-negotiable expectations, not aspirational footnotes.

AI in Oncology

This week's AI in Oncology section spans a wide spectrum of maturity and methodological rigour, collectively illustrating both the promise and persistent gaps in clinical AI translation. The most significant contribution is the externally validated mammography-derived AI risk model (Paper 5), which achieved a consistent 10-year AUC of 0.72 across US and Swedish cohorts and identified 33% of incident breast cancers in the top-decile risk group — outperforming Tyrer-Cuzick, BCSC, and Mirai. This is the first long-term image-derived AI model validated for primary prevention at this scale, and it sets a new benchmark for the field. The scoping review of explainable AI in cancer imaging (Paper 4) provides essential context: across 371 studies from 2017–2024, only 5.2% quantitatively validated their explanations, 17.5% shared code, and just 12.1% reported any clinical decision support integration. Deep learning dominates, but post hoc saliency maps remain the interpretability workhorse — a gap that complicates regulatory and clinical trust. The ccRCC aging-related gene model (Paper 1) applies unspecified ML algorithms to TCGA data, yielding a seven-gene prognostic signature with immunotherapy implications — but experimental validation was incomplete for two key genes (FOXM1, CDKN2A), tempering confidence. The PG-DLBCL nomogram (Paper 2) was misclassified as an AI paper; Cox regression, not machine learning, underpins its predictions, though its prognostic granularity for a rare entity is clinically useful. The LLM comparison study (Paper 3) evaluates DeepSeek versus Kimi for urological patient education, including bladder cancer — finding modest, inconsistent advantages for each — a timely but preliminary contribution. Taken together, the section highlights a recurring theme: model development is outpacing validation and deployment infrastructure across virtually every AI oncology subfield.

PUBMED  ·  2026-05-15  ·  Significance: 2/5

This bioinformatics and experimental study investigated the role of aging-related genes (ARGs) in clear cell renal cell carcinoma (ccRCC) prognosis. Using transcriptomic data from the TCGA-ccRCC cohort, the authors intersected differentially expressed genes with known ARGs to identify 25 candidate genes, then applied machine learning algorithms to select seven with prognostic relevance: PCK1 (protective) and TOP2A, TFAP2A, CCNA2, FOXM1, CDKN2A, and PLAU (risk-associated). A risk stratification model was constructed and validated via survival analysis, with high-risk patients showing reduced overall survival. Immune infiltration and checkpoint expression differed between risk groups. Drug prediction identified 69 potential therapeutic compounds including tyrosine kinase and mTOR inhibitors. RT-qPCR validated expression of five genes in tissue samples, though discrepancies were noted for FOXM1 and CDKN2A, limiting full experimental confirmation.

Key Findings

PUBMED  ·  2026-05-15  ·  Significance: 3/5

This retrospective study used data from the SEER (Surveillance, Epidemiology, and End Results) database to develop and validate prognostic nomograms for predicting overall survival (OS) and cancer-specific survival (CSS) in patients with primary glandular diffuse large B-cell lymphoma (PG-DLBCL) — a rare malignancy affecting four glandular sites. Patients were split 70/30 into training and validation cohorts, and Cox regression-based nomograms were constructed. Key prognostic factors identified included age, Ann Arbor stage, tumor site, and treatment modalities. Survival outcomes varied markedly by glandular site: 5-year OS was 38.10% for primary adrenal DLBCL versus 72.79% for primary thyroid DLBCL. The nomograms demonstrated moderate discriminative performance, with C-indices of 0.75 for OS and 0.77 for CSS.

Key Findings

PUBMED  ·  2026-05-16  ·  Significance: 2/5

This study evaluates and compares two Chinese large language models (LLMs) — DeepSeek and Kimi — for their efficacy in providing health management information across six urological and andrological diseases. Thirty questions were assessed using quality evaluation tools (PEMAT-AI, DISCERN-AI), expert assessments of accuracy and safety, and readability metrics. DeepSeek demonstrated superior quality scores in responses to questions about kidney stone surgical options and bladder cancer treatment modalities, and outperformed Kimi in accuracy and safety specifically for dietary advice on kidney stones. Kimi, however, performed better on readability metrics overall. Neither model showed statistically significant differences across most disease-specific categories. The study concludes that both models have clinical applicability but require targeted improvements in complementary domains.

Key Findings

PUBMED  ·  2026-05-20  ·  Significance: 2/5

This scoping review systematically maps the application of explainable artificial intelligence (xAI) methods across radiologic cancer imaging, analysing 371 peer-reviewed studies published between 2017 and 2024 (PubMed and Scopus search). The review characterises xAI methodology, imaging modalities, validation approaches, and clinical integration status. Breast, lung, and brain cancers dominated the literature, with CT and MRI as the primary modalities. Deep learning underpinned 70% of studies, and post hoc visualisation methods (e.g., saliency maps) were by far the most common explainability approach. The review identifies critical gaps: quantitative validation of explanations remains rare, only 17.5% of studies shared code, and just 12.1% reported any decision support system integration. The authors conclude that the field lacks standardisation and that most xAI models remain far from clinical deployment.

Key Findings

PUBMED  ·  2026-05-20  ·  Significance: 4/5

A long-term AI-based breast cancer risk model, developed using mammography image-derived features, was externally validated across three diverse population cohorts in the US and Sweden (Olmsted County, KARMA, and EMBED), totalling 8,696 individuals with a median 10-year follow-up. Note: despite the provided glossary suggesting AI = aromatase inhibitor, context clearly indicates AI = Artificial Intelligence throughout this paper. The model estimated absolute 10-year breast cancer risk at study entry and was compared against established clinical tools (Tyrer-Cuzick-v8, BCSC-v3) and the image-based Mirai tool. The AI model achieved a 10-year time-dependent AUC of 0.72 in both Olmsted and KARMA cohorts, with well-calibrated expected-to-observed ratios of 0.99. It identified 33% of breast cancers in the top 10% highest-risk individuals, significantly outperforming all comparator tools.

Key Findings