Deep learning-driven glioblastoma diagnosis from histopathological images via single-cell segmentation and morphological analysis

Abstract Glioblastoma (GBM) exhibits a high recurrence rate of 95% due to its highly infiltrative nature and marked heterogeneity, making its diagnosis challenging. These characteristics complicate the standardization of diagnostic criteria and contribute to significant interpathologist variability. In response, artificial intelligence-based tools are increasingly being developed to support the diagnostic process. This study presents a complete processing pipeline that integrates deep learning and machine learning techniques to facilitate the screening of hematoxylin and eosin-stained histopathological images, aiming to reduce the workload in pathology departments. The proposed workflow includes a convolutional neural network capable of detecting and segmenting individual cells, achieving 89.5% pixel-wise precision and a DICE coefficient (DSC) of 88.4%, indicating a high degree of spatial overlap between the predicted and reference regions. Following segmentation, a set of morphological features is extracted, including cell area, circularity, cell count, and the average distance between neighboring cells. The extracted features were subsequently used for tissue classification into tumorous and non-tumorous categories via a random forest algorithm, resulting in a high classification accuracy of 94.15%, thereby suggesting the robustness and reliability of the proposed approach. The applied methodology was validated on a cohort of 82 patients, comprising both tumor and healthy tissue, with a total of 1557 images and more than 4 million annotated cells.