Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Pediatric brain tumor classification using digital histopathology and deep learning: evaluation of SOTA methods on a multi-center Swedish cohort

Created by
  • Haebom

Author

Iulian Emil Tampu, Per Nyman, Christoforos Spyretos, Ida Blystad, Alia Shamikh, Gabriela Prochazka, Teresita D iaz de St{\aa}hl, Johanna Sandgren, Peter Lundberg, Neda Haj-Hosseini

Outline

To evaluate the potential of computational pathology methods for diagnosing pediatric brain tumors, we implemented a weakly supervised multiple-instance learning (MIL) method using brain tumor tissue slide images (WSI) from 540 pediatric patients (mean age 8.5 years) collected from a multicenter cohort at six university hospitals in Sweden. Three pretrained feature extractors, ResNet50, UNI, and CONCH, were used to extract patch-level features from the WSI, and ABMIL or CLAM were used to aggregate features for patient-level classification. Models were evaluated on three classification tasks—type, family, and type—based on a hierarchical classification of pediatric brain tumors. The models were trained on data from two centers and tested on data from the remaining four centers to assess their generalization performance. The highest classification performance was achieved using UNI features and ABMIL aggregation, with Matthew correlation coefficients of 0.76±0.04, 0.63±0.04, and 0.60±0.05 for type, family, and type classification, respectively. Models using UNI and CONCH features had better generalization performance than models using ResNet50.

Takeaways, Limitations

•
Takeaways:
◦
State-of-the-art computational pathology methods demonstrate effectiveness in diagnosing pediatric brain tumors at multiple hierarchical levels.
◦
It also shows significant generalization performance on multi-institutional datasets.
◦
Validation of the superior performance of the UNI feature extractor and the ABMIL aggregation method.
•
Limitations:
◦
Generalization performance on multicenter datasets still has room for improvement (there are performance differences across centers).
◦
The size of the dataset used in the study may be relatively small compared to larger datasets.
◦
There may be dependencies on specific feature extractors and MIL methods.
👍
Made with Slashpage