Fusion of histopathology images and tabular data for pediatric brain tumor classification

Sammanfattning

This study investigates whether integrating tabular clinical data with histopathological whole slide image (WSI) features improves classification performance for pediatric central nervous system (CNS) tumors, using the Children’s Brain Tumor Network (CBTN) dataset. Five tumor types are classified, including low-grade glioma (LGG), high-grade glioma (HGG), medulloblastoma (MB), ependymoma (EP), and ganglioglioma (GG), spanning 513 patients and 1,054 hematoxylin and eosin (H&E)-stained WSIs. WSI features were extracted using the contrastive learning from captions for histopathology (CONCH) foundation model and classified with the clustering- constrained-attention multiple-instance learning (CLAM) framework. Three tabular variables were considered, patient age at diagnosis, biological sex, and Ki-67 proliferation index. Six multimodal fusion strategies were evaluated, comprising three early fusion methods (concatenation, cross-gate, and FiLM) and three late fusion methods (probabil- ity averaging, weighted averaging, and meta-learner stacking), each tested across seven tabular feature combinations (age, sex, Ki-67, age + sex, age + Ki-67, sex + Ki-67, and age + sex + Ki-67), resulting in 42 configurations in total. The image-only baseline achieved a balanced accuracy of 0.76±0.05. The tabular-only baseline performed substantially lower at 0.45±0.05, suggesting that clinical tabular variables alone are insufficient for reliable classification. Both early and late fusion improved over the image-only baseline, with five out of six configurations achieving statistically significant improvements after Bonferroni correction. The best early fusion configuration using concatenation of H&E WSI with age + sex + Ki-67 achieved a balanced accuracy of 0.80 ± 0.05. The best late fusion configuration using meta-learner stacking of H&E WSI with Ki-67 achieved 0.78 ± 0.05. Early fusion outperformed late fusion in terms of balanced accuracy, and was statistically superior in 5 out of 9 pairwise comparisons (three early fusion strategies each compared against three late fusion strategies) after Bonferroni correction. Values are rounded to two decimal places for readability. In conclusion, Ki-67 was the most informative tabular feature across fusion configurations. Age and sex alone did not consistently improve over the image-only baseline, but in early fusion, combining them with Ki-67 yielded higher balanced accuracy than using Ki-67 alone. Overall, the fusion of age, sex, and Ki-67 with H&E WSI features improves five-class pediatric brain tumor classification, and the choice of fusion strategy is important.