Tumors of the kidney comprise several subtypes with distinct genetic profiles, treatment requirements, and prognoses, including clear cell renal cell carcinoma (ccRCC), papillary RCC (pRCC), chromophobe RCC (chRCC), and renal oncocytoma (RO). Accurate subtype classification is essential because the treatment approaches differ substantially across these groups, and an incorrect diagnosis can lead to under- or over-treatment.
A particularly difficult diagnostic challenge involves distinguishing chromophobe RCC from renal oncocytoma, two tumor types that can appear histologically similar under the microscope. While renal oncocytoma is benign, chromophobe RCC has malignant potential with risk of metastasis. Standard morphology-based pathology struggles to reliably separate them, and current immunohistochemical markers have been shown to be unreliable and lack standardization.
MALDI mass spectrometry imaging (MSI) combines spatial metabolomics and conventional histology by generating in situ chemical maps of tissue sections, making it well-suited for multiparametric, multi-omics analysis. Although MSI simultaneously generates both metabolomic and morphometric data, the morphometric component had never previously been exploited alongside metabolomics to improve machine learning-based tumor classification.
The study was conducted on 853 patients from multiple centers, representing the largest cohort used for MALDI MSI-based renal tumor classification at the time of publication. The cohort included ccRCC (n = 552), pRCC (n = 122), chRCC (n = 108), and renal oncocytoma (n = 71), all from formalin-fixed paraffin-embedded (FFPE) tissue samples, the standard tissue preservation format used in routine pathology.
MALDI MSI was performed on tissue sections to generate both spatial metabolomic profiles and morphometric image data from the same tissue section after hematoxylin and eosin (H&E) staining. The metabolomic feature set comprised 2,111 metabolite ion signals, while the morphometric feature set included 110 features describing tissue and cell compartment color, shape, and size. Feature selection used Kruskal-Wallis testing on the training set.
Three separate random forest classifiers were trained: one on morphometric data alone, one on metabolomic data alone, and one on the combined dataset. Patients were randomly split into a training set (two-thirds) and an independent validation set (one-third), with data normalized separately for each set. This design allowed direct comparison of the three approaches on the same held-out test data, with the synergistic classifier directly answering whether combining data types adds value beyond either alone.
The morphometric classifier achieved a mean accuracy of 77.81% across all four subtypes, performing best for ccRCC (F1-score 86.22%) but poorly for chRCC (F1-score 52.17%). The metabolomics classifier improved overall accuracy to 85.13%, with substantially better chRCC performance (F1-score 76.46%) and strong ccRCC classification (F1-score 91.06%). Only for pRCC did the metabolomics classifier perform slightly worse than morphometrics (63.66% versus 67.56%).
The combined classifier trained on both morphometric and metabolomic features achieved a mean accuracy of 88.04%, outperforming both individual classifiers for every tumor subtype. F1-scores for the combined classifier were 92.54% for ccRCC, 76.73% for pRCC, 77.15% for chRCC, and 84.65% for RO. The complementarity of the two data types was particularly evident for pRCC, where the combined classifier improved performance by up to 10 percentage points over the metabolomics-only classifier.
Feature importance analysis using Gini importance showed that the top 50 features in the combined classifier represented an even mixture of metabolomic and morphometric features. The four most important individual features were metabolites, but morphometric features appeared throughout the top 50, and the combined top-feature list was more informative than either data type alone could provide.
Because chromophobe RCC and renal oncocytoma are the two subtypes most difficult to distinguish histologically (Figure 3A shows their similar H&E appearance), the three classifiers were also evaluated on this two-class problem in isolation. This targeted analysis reflects the most clinically important diagnostic challenge in renal tumor pathology.
On morphometric data alone, the two-class classifier achieved 81.27% accuracy. On metabolomics data, accuracy rose to 89.49%. The combined classifier reached 91.0% accuracy, with F1-scores of 92.70% for chRCC and 88.07% for RO, confirming that combining data types provides meaningful benefit even for the most challenging subtype pair.
For the chRCC-versus-RO binary problem, fewer morphometric features (40%) appeared in the top 50 compared to the four-class problem, reflecting that metabolomics data carries more discriminatory power when morphological differences between the two subtypes are subtle. Nevertheless, morphometric features continued to be ranked among the top 50 in the combined model, demonstrating that they add complementary information even when metabolomics features are dominant.
MALDI (matrix-assisted laser desorption/ionization) mass spectrometry imaging generates spatially resolved chemical profiles of tissue sections by desorbing and ionizing molecules from discrete tissue spots and measuring their mass-to-charge ratios. The result is a spatial map of hundreds to thousands of metabolite signals across the tissue, providing information about the chemical composition of different tissue regions that cannot be obtained from standard histology.
A key practical advantage of MALDI MSI is that it is compatible with FFPE tissue, the standard preservation method used in hospital pathology archives. This compatibility means the technology can be applied retrospectively to existing tissue archives, allowing large-scale studies without requiring fresh frozen specimens. FFPE compatibility is a major step toward clinical translation of MALDI-based diagnostics.
The morphometric data derived from H&E staining is inherently co-registered with the MALDI MSI metabolomic data, since both are acquired from the same tissue section. This perfect spatial alignment means that morphometric and metabolomic features can be combined without any registration errors, making their integration computationally straightforward and biologically meaningful as they describe the same tissue regions.
This study demonstrates for the first time that morphometric features derived from H&E-stained tissue sections can systematically improve the predictive accuracy of metabolomics-based machine learning classifiers when combined. The improvement was consistent across all subtypes and was largest for the pRCC subtype, where morphometric information compensated for gaps in metabolomic discrimination.
The authors propose that this synergistic multi-modal approach should be applied broadly to other cancer types and diagnostic challenges where MALDI MSI is used, since morphometric data are already inherently generated during every MSI experiment but have historically been discarded. Leveraging this previously unused data stream costs nothing additional and consistently improves classification performance.
Future work should focus on prospective clinical validation, reproducibility across different MALDI instrument platforms and tissue preparation protocols, and integration with other molecular data types such as proteomics and genomics. If the 91% accuracy for chRCC versus RO is confirmed prospectively, this approach could meaningfully reduce the need for additional molecular tests like immunohistochemistry and genetic profiling in the most diagnostically challenging kidney tumor cases.