Background Foundation models have shown promise in medical AI by learning flexible features from large datasets, offering new opportunities for improving endpoint prediction. However, usage of foundation models for endpoint prediction using routine imaging in head and neck squamous cell carcinoma patients remains unexplored. Within this study, we evaluated the potential of foundation-model based multiple instance learning for prediction of 2-year overall survival, locoregional control and freedom from distant metastasis across three external head and neck squamous cell carcinoma patient cohorts using 2D, multiview and 3D approaches while comparing prediction and stratification performance with handcrafted radiomics and clinical baselines. Results 2D multiple-instance learning models achieved 2-year test area under the receiver-operator curve (AUROC) range of 0.75–0.84 for 2-year overall survival, 0.66–0.75 for 2-year locoregional control and 0.71–0.78 for 2-year freedom from distant metastasis across three different external cohorts, outperforming multiview and 3D multiple instance learning models (AUROC range: 0.50–0.77, p <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44398_2025_3_Article_IEq1.gif" Format="GIF" Height="15" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$\ge$</EquationSource> <EquationSource Format="MATHML"><math> <mo>≥</mo> </math></EquationSource> </InlineEquation> 0.15) and showing comparable or superior performance to handcrafted radiomics (AUROC range: 0.64–0.74, p <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44398_2025_3_Article_IEq1.gif" Format="GIF" Height="15" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$\ge$</EquationSource> <EquationSource Format="MATHML"><math> <mo>≥</mo> </math></EquationSource> </InlineEquation> 0.012). Significant stratification was observed from the 2D MIL models (hazard ratios: 2.14–4.77, p <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44398_2025_3_Article_IEq3.gif" Format="GIF" Height="15" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$\le$</EquationSource> <EquationSource Format="MATHML"><math> <mo>≤</mo> </math></EquationSource> </InlineEquation> 0.039). 2D MIL models were also shown to learn endpoint-specific correlation patterns such as N-stage for 2-year freedom from distant metastasis prognosis. Multimodal enhancement of 2-year OS/FFDM (AUROC range: 0.82–0.87, p <InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="44398_2025_3_Article_IEq3.gif" Format="GIF" Height="15" Rendition="HTML" Resolution="72" Type="Linedraw" Width="19" /> </InlineMediaObject> <EquationSource Format="TEX">$\le$</EquationSource> <EquationSource Format="MATHML"><math> <mo>≤</mo> </math></EquationSource> </InlineEquation> 0.018) for patients without human papilloma virus positive tumors. Conclusions FM-based 2D MIL demonstrates promise in HNSCC risk prediction as well as stratification of clinical outcomes. The models match or outperform radiomics baselines, learning clinically-related patterns and showing enhancement of clinical baselines in non-human papilloma virus positive patients.

End-to-end prediction of clinical outcomes in head and neck squamous cell carcinoma with foundation model-based multiple instance learning

期刊论文
出版模式：

开放获取
发表日期：

2025年6月24日

查看全文

Asier Rabasco Meneghetti,
Marta Ligero Hernández,
Jens-Peter Kühn,
Steffen Löck,
Zunamys Itzell Carrero,
Raquel Perez-Lopez,
Keno K. Bressem,
Titus J. Brinker,
Alexander T. Pearson,
Daniel Truhn,
Sven Nebelung,
Jakob Nikolas Kather

摘要

Background

Foundation models have shown promise in medical AI by learning flexible features from large datasets, offering new opportunities for improving endpoint prediction. However, usage of foundation models for endpoint prediction using routine imaging in head and neck squamous cell carcinoma patients remains unexplored. Within this study, we evaluated the potential of foundation-model based multiple instance learning for prediction of 2-year overall survival, locoregional control and freedom from distant metastasis across three external head and neck squamous cell carcinoma patient cohorts using 2D, multiview and 3D approaches while comparing prediction and stratification performance with handcrafted radiomics and clinical baselines.

Results

2D multiple-instance learning models achieved 2-year test area under the receiver-operator curve (AUROC) range of 0.75–0.84 for 2-year overall survival, 0.66–0.75 for 2-year locoregional control and 0.71–0.78 for 2-year freedom from distant metastasis across three different external cohorts, outperforming multiview and 3D multiple instance learning models (AUROC range: 0.50–0.77, p $\ge$ $\geq$ 0.15) and showing comparable or superior performance to handcrafted radiomics (AUROC range: 0.64–0.74, p $\ge$ $\geq$ 0.012). Significant stratification was observed from the 2D MIL models (hazard ratios: 2.14–4.77, p $\le$ $\leq$ 0.039). 2D MIL models were also shown to learn endpoint-specific correlation patterns such as N-stage for 2-year freedom from distant metastasis prognosis. Multimodal enhancement of 2-year OS/FFDM (AUROC range: 0.82–0.87, p $\le$ $\leq$ 0.018) for patients without human papilloma virus positive tumors.

Conclusions

FM-based 2D MIL demonstrates promise in HNSCC risk prediction as well as stratification of clinical outcomes. The models match or outperform radiomics baselines, learning clinically-related patterns and showing enhancement of clinical baselines in non-human papilloma virus positive patients.