Transforming Real-Time Diagnosis of Brain Tumors: AI System Identifies Gliomas from Cryosection Images

Transforming Brain Tumor Diagnosis: AI System Identifies Gliomas from Cryosection Images | The Lifesciences Magazine

A recent study conducted by researchers from Harvard Medical School, the University of Pennsylvania’s Perelman School of Medicine, and Cedars-Sinai Medical Center has introduced a machine-learning approach for cryosection pathology. This innovative model enables the accurate classification of gliomas using digital images of cryosection samples during surgery, showcasing the potential of artificial intelligence in facilitating real-time intraoperative Brain Tumor Diagnosis.

Dr. Kun-Hsing “Kun” Yu, MD, PhD, an assistant professor at Harvard Medical School, led this research effort, addressing the clinical challenge of achieving precise brain cancer diagnoses during surgery. The AI models developed in this study offer real-time brain cancer diagnoses and predictions regarding molecular subtypes, providing invaluable support to neurosurgeons when determining treatment strategies.

The findings were published in the journal “Med.”

The study sought to overcome the limitations of cryosection pathology, a critical process for neurosurgeons to rapidly and accurately assess tissue samples during brain tumor removal surgeries. Cryosectioning involves freezing tissue samples and rapidly evaluating them under a microscope to guide the surgical process.

Dr. Yu explained that neurosurgeons often send brain cancer samples to pathologists during surgery, who then utilize cryosection techniques to process and evaluate the tissue within 10-15 minutes. While cryosectioning is faster than traditional histology methods, the resulting image quality is lower, leading to potential diagnostic challenges and the inability to determine specific diagnostic categories that now require genetic data.

To address these challenges, the researchers created a machine learning model called “Cryosection Histopathology Assessment and Review Machine” (CHARM). CHARM is a context-aware machine learning approach that uses a vision transformer to diagnose gliomas during surgery. It employs weakly supervised learning to analyze whole-slide images and provide slide-level labels for various diagnostic criteria. The model assigns separate predictions to each image tile, aggregating these predictions to provide a patient-level Brain Tumor Diagnosis.

CHARM excels by integrating diagnostic signals from cancer cells and their surrounding tissues, which is a significant advantage when dealing with low-quality samples from different hospitals.

Brain Tumor Diagnosis – Personalized Medicine for Gliomas

CHARM was evaluated using samples from over 1,500 glioma patients, demonstrating remarkable performance. It achieved an area under the curve (AUC) value of 0.98 for identifying malignant cells, outperforming conventional convolutional neural networks (CNNs) with an AUC of 0.88. The model accurately classified histologic grades and glioma subtypes in line with the 2021 WHO guidelines.

CHARM’s quick diagnostic capabilities offer substantial benefits, as it can complete its analysis in less than a second compared to the several days required by standard molecular profiling methods. Additionally, it accurately predicts genetic alterations related to IDH mutations, ATRX, and TP53, contributing to more efficient glioma Brain Tumor diagnosis during surgery.

A current limitation of the model is the need for standard tissue processing and digitization of cryosection images. However, emerging non-invasive tissue analysis methods could enhance CHARM’s capabilities in the future.

The authors of the study have made their models and codes freely available for research in healthcare centers worldwide. They are also actively exploring the application of their approach to other types of brain cancers and resource-limited settings.

Also Read: Targeted Therapy Vorasidenib Offers Hope for Deadly Brain Tumors

Share Now