Neuro Oncol. 2026 Nov 12. pii: noag283. [Epub ahead of print]
Siddhesh P Thakur,
Carla Pitarch-Abaigar,
Sylwia Malec,
Akis Linardos,
Shubham Innani,
Sanyukta Adap,
Ujjwal Baid,
Shahriar Faghani,
Mana Moassefi,
MacLean P Nasrallah,
Jared T Ahrendsen,
Valeria Barresi,
Maria A Gubbiotti,
Giselle Y Lopez,
Calixto-Hope G Lucas,
Michael L Miller,
Dominic LaBella,
Marek Wodzinski,
Henning Müller,
Juexin Zhang,
Ying Weng,
Ke Chen,
Rithik Kumar,
Vinit,
Madhav Arora,
Twinkle Dhingra,
Syed Rameem Zahra,
Vansh Kakkar,
Ankur Gupta,
Alvaro A Sandino,
Olivia Krebs,
Jeffrey Helgager,
David S Becerra,
Nelly J Sandino,
Eduardo Romero,
Pallavi Tiwari,
Bhakti Baheti,
Verena Chung,
Sarthak Pati,
Alexandros Karargyris,
Hasan Kassem,
Matthew Pease,
Benedikt Wiestler,
Bjoern Menze,
Maria Correia de Verdier,
Raymond Huang,
Jeff Rudie,
Zhifan Jiang,
Marius George Linguraru,
Lee A D Cooper,
Martin van den Bent,
Timothy Cloughesy,
Michael A Vogelbaum,
Patrick Y Wen,
Susan Chang,
William R Bell,
Jason T Huse,
Spyridon Bakas.
BACKGROUND: Automated classification of glioblastoma (IDH-wt, WHO CNS Gr.4) histopathologic sub-regions in whole slide images (WSI) is essential for quantitative tumor characterization, but faces substantial challenges from intra-tumoral heterogeneity and inter-observer annotation variability. The AI-RANO & RANO-Pathology working groups organized BraTS-Path 2024, representing the first international benchmarking environment and publicly available spatially annotated dataset, towards promoting the development and fair evaluation of Artificial Intelligence (AI) models to classify specific sub-regions: cellular tumor, pseudopalisading necrosis (PN), microvascular proliferation, geographic necrosis, infiltration into the cortex, and white matter penetration (WM).
METHODS: Multi-institutional H&E-stained WSIs had reference standard samples generated by eight expert neuropathologists, following a rigorous peer review protocol, resulting in 309,118 data samples (training/validation/hold-out testing: 96,485/62,387/150,246). Quantitative performance evaluation of six AI approaches was conducted based on multiple metrics, including Accuracy and Matthew's correlation coefficient (MCC). A robust statistically-aware ranking strategy offered statistical significance of the final rankings.
RESULTS: Top-performing approaches employed common strategies, including large pre-trained models, ensembling, and extensive data augmentation, reporting Accuracy=0.69 and MCC=0.58. Key challenges included severe class imbalance, hampering minority class performance, and inter-annotator variability, especially for ambiguous sub-regions such as PN or WM.
CONCLUSIONS: BraTS-Path represents a valuable public resource for glioblastoma histopathologic sub-region assessment. Our findings indicate the specific task remaining an open research problem, highlighting the influence of rare sub-regions, the absence of a consistent "gold standard" annotation protocol, and the potential of foundation models to improve generalization, guiding future research toward developing more reliable AI tools for quantitative neuropathology.
Keywords: AI; RANO; benchmark; glioblastoma; pathology