408P Machine learning for glioblastoma screening from histopathology whole slide imaging

Background: Glioblastoma (GBM) is one of the most common malignant primary brain tumours, which accounts for 60-70% of all glioma. Conventional diagnosis and post-operation treatment plan for glioblastoma is mainly based on the identification of tumor cells on haematoxylin and eosin stained (H&E) histopathological slides by both experienced medical technologist and pathologist. The recent development of digital whole slide scanners allows the histopathological image analysis feasible. This study aimed to build an image feature based model using histopathology whole slide images to differentiate patients with GBM from healthy control (HC).

Methods: 628H&E images were collected from 30 GBM patients, obtained from public database the Cancer Image Archive. Another 554 images were collected from 30HC. 22 Gray level Co-occurrence Matrix (GLCM) image features were retrieved from the collected images by an in-house developed software. Both the GBM and HC groups were divided into a training and a validation group on 70:30 basis. A random forest algorithm was used to build the model using training set, and validated by the validation set by R software.

Results: The receiver operating characteristics (ROC) curve showed that the average class and overall accuracy achieved 94.6%. The area under the curve (AUC) was 0.98, with sensitivity of94.1% and sensitivity of 95.1%.

Conclusions: The machine learning model can accurately differentiate GBM from HC in histopathology whole slide imaging. The application can be fully automatic and it is practical in applying for GBM screening.