Combing a Deep Learning-Based Radiomics With Liquid Biopsy for Preoperative and Non-invasive Diagnosis of Glioma

Gliomas are the most common primary intracranial malignancies, accounting for 27% of all primary brain tumors, and approximately 100,000 people are diagnosed with diffuse gliomas worldwide each year. To date, "integrated diagnosis" was considered the gold standard for glioma diagnosis, which combines histopathology, molecular pathology, and World Health Organization (WHO) grade. Previous glioma diagnostic criteria have primarily relied on histopathological biopsies, while histological classification has traditionally been determined based on tumor morphology, resulting in intra-observer variability due to intra-tumor spatial heterogeneity and sampling errors. In addition, Traditional histopathology is somewhat difficult to explain why patients with the same pathology have significantly different survival. Over the past decade, advances in molecular pathology and histopathology detection techniques have deepened our understanding in the molecular features and biology of gliomas. Increasing evidence revealed the important role of molecular status in the " integrated diagnosis" of glioma. In particular, after the concept of molecular diagnosis was proposed by the 2016 WHO Central Nervous System (CNS) classification, the 2021 CNS classification (CNS5) reemphasized the importance of several molecular biomarkers in gliomas diagnosis and treatment guidance, including isocitrate dehydrogenase gene (IDH ) mutation status, alpha-thalassemia/mental retardation syndrome X-linked (ATRX) deletion status, 1p19q deletion status, telomerase reverse transcriptase (TERT) promoter mutation, etc. Which the objective is to classify the tumor subtypes more systematically, and categorize the glioma patients with similar efficacy and prognosis into a subgroup.

The current standard of therapy for gliomas is surgical resection followed by radiotherapy and/or chemotherapy based on clinical and tumor grade and molecular characteristics. Preoperatively non-invasive and accurate early " integrated diagnosis" will bring great benefits to the treatment and prognosis of patients, especially for those with special tumor location that cannot receive craniotomy or needle biopsy. Such special patients can take experimental radiotherapy and chemotherapy based on non-invasive diagnosis results. Although diagnostic criteria for molecular information in gliomas are often based on tissue biopsy, other techniques, such as radiomics, radiogenomics, and liquid biopsy, have shown promise. At present, conventional magnetic resonance imaging (MRI) scans are still the main method to assist in the diagnosis of gliomas, including pre- and post- contrast T1w, T2w, and T2w-FLAIR. Multimodal radiomics based on deep learning (DL) can analyze patterns of intratumor heterogeneity and tumor imaging features that are imperceptible by the human eye, so as to conduct " integrated prediction" of gliomas18. Up to now, most studies have focused on using ML algorithms to construct novel radiomic model to predict glioma, R van der Voort et al. developed the multi-task conventional neural network (CNN) model and achieved a glioma grade (II/III/IV) with AUC of 0.81, IDH-AUC of 90%, 1p19q co-deletion AUC of 0.85 in the test set. The best DL model developed by Matsui et al. achieved an overall accuracy of 65.9% in predicting IDH mutation and 1p/19q co-deletion. Also, the multi-task CNN model constructed by Decuyper et al. achieved 94%, 86%, and 87% accuracy in predicting grades, IDH mutations, and 1p/19q co-deletion states in external validation. The model constructed by Luo et al. achieved 83.9% and 80.4% in external tests for histological and molecular subtype diagnosis. In addition to the "integrated prediction", there exists many models that only predicting glioma grading or single molecular markers. Meanwhile, previous studies were based on the 2016 CNS classification for glioma grading and molecular subtypes prediction. Therefore, a multi-task DL radiomics model for preoperatively and non-invasively predicting glioma grading and more extensive molecular markers is urgently needed according to the latest 2021 CNS classification.

Although radiomics has showed some feasibility in predicting tumor molecular pathology, it is ridiculous to administer precision targeted therapy solely on the basis of this prediction. Therefore, we hope to provide more clinical evidence for the molecular pathological diagnosis of gliomas patients by using liquid biopsy technique as an important complement of radiomics. Circulating tumor cell (CTC), as one of the liquid biopsy techniques, shares the same final objective to preoperatively non-invasive and accurate diagnosis of gliomas.

Based on the several limitations of the current diagnostic models of glioma, and the combined methods of radiomics and liquid biopsy have great potential to non-invasive diagnose glioma grading and molecular markers since they are both easy to perform. Furthermore, to our knowledge, there has been no study of preoperative non-invasive diagnosis of glioma in the context of liquid biopsy-assisted radiomics.

Therefore, this study has the following objectives. First, according to the guidance of 2021 WHO of CNS classification, we constructed and externally tested a multi-task DL model for simultaneous diagnosis of tumor segmentation, glioma classification and more extensive molecular subtype, including IDH mutation, ATRX deletion status, 1p19q co-deletion, TERT gene mutation status, etc. Second, based on the same ultimate purpose of liquid biopsy and radiomics, we innovatively put forward the concept and idea of combining radiomics and liquid biopsy technology to improve the diagnosis of glioma. And through our study, it will provide some clinical validation for this concept, hoping to supply some new ideas for subsequent research and supporting clinical decision-making.