Updates on glioblastoma multiforme: From epidemiology to imaging and artificial intelligence

Abstract

Glioblastoma multiforme (GBM) are the most aggressive and common tumors in the central nervous system. GBM are classified as grade IV according to the World Health Organization. The incidence of GBM slightly differs among countries. The etiology of GBM has not been entirely clarified. No risk factors such as smoking, chemicals or dietary can be identified for GBM. Only the exposure to high radiation dose such as radiotherapy of head and neck cancers have been reported to increase the risk of glioma tumors. In this review, the authors attempted to cover several aspects of GBM. This review was based on a collection of recent publications from different research fields but all related to GBM in order to shed the light on this disease. We highlighted the current insights of GBM in the aspects of epidemiology, pathogenesis, etiology, molecular genetics, imaging technologies, artificial intelligence and treatment. A literature review was conducted for GBM with relevant keywords. Although GBM was known since several decades, its causes are still confounding, and its early detection is often unpredictable. Since the hereditary aspect of GBM is very low, there remains as the common symptoms the interference with normal brain function, memory loss, unusual behavior, headaches and seizures. The progress in GBM treatment is not satisfactory even with the deployment of huge efforts and financial costs in many domains like gene therapy, surgery and chemoradiotherapy. Despite the rapid developments of the standard treatment for GBM, the trend of survival rate did not change among years.

Key Words: Glioblastoma; Cancer; Radiation therapy; Chemotherapy; Imaging; Positron emission tomography; Magnetic resonance imaging; Artificial intelligence

Core Tip: Glioblastoma remains one of the most lethal brain cancers due to its complexity and resistance to therapy. Artificial intelligence (AI) offers transformative potential in early detection, precision surgery, personalized treatment, and drug discovery. By integrating imaging, genomics, and clinical data, AI paves the way to improved capabilities for faster and more effective glioblastoma care.

INTRODUCTION

Glioma is a term used to describe the brain tumors which originate from glial supportive cells in the brain or spinal cord. Three types of gliomas are identified based on their cell type: Astrocytoma, ependymomas and oligodendrogliomas[1]. Glioblastoma multiforme (GBM) is the most severe type of astrocytoma. GBM are the most aggressive and common tumors in the central nervous system (CNS). These tumors can affect children; however, their incidence rate is increased with patient age[2]. No specific causes have been known for GBM[3]. The heterogeneity of these tumors makes them the most challenging cancer to treat[4].

The World Health Organization (WHO) proposed classification grades for brain tumors. It ranges from grade I to grade IV with the highest malignancy given to grade IV[5]. The classification was based on histological similarities of morphological cells and their levels of differentiation. In 2016, the WHO published the new classification system of the central nervous tumors[6]. The new system adds a new genetic and molecular information that facilitates the tumor classification. New terminology was introduced for GBM to reflect the genetic features associated with it, such as the mutation of isocitrate dehydrogenase (IDH) gene. Besides, epithelioid glioblastoma was added to the classification. Epithelioid GBMs are rare tumors, were classified earlier as secondary glioblastoma, which accounts for 10% of GBM cases[7]. These epithelioid tumors found mainly in younger patients and were associated with poor prognosis. The poor prognosis was explained as an increased capacity for leptomeningeal dissemination[8]. In summary, the new classification of GBM includes GBM-IDH wild-type (Giant cell GBM, Gliosarcoma, and Epithelioid GBM), GBM-IDH mutant and GBM-NOS (not otherwise specified). The GBM tumors are characterized by tumor cell infiltration, pleomorphic cells, extremely rapid growth of intravascular microthrombi, hypoxia, and necrosis[9].

The symptoms of GBM depends on tumor size and location. The common signs of documented symptoms for GBM are headache, nausea, vomiting, confusion, memory loss, personality changes, balance difficulty, vision problems, speech difficulties, and seizures[10].

The steps of GBM diagnosis starts with a neurological test to examine the patient's senses (vision and hearing), balance, strength, and reflexes. In case of clues of GBM with the neurological exam, some imaging exams are used to diagnose brain tumors such as magnetic resonance imaging (MRI), computerized tomography (CT), and positron emission tomography (PET). Besides, genetic tests can be used to determine the metastatic of GBM and the histological properties through a surgical biopsy of the tumor[11].

The current techniques of GBM treatment include surgery, chemotherapy, and radiotherapy[12]. Due to the poor prognosis with GBM, the survival rate for 5 years is about 5%[13]. Patients usually survive for about 15 months after utilizing all possible cancer treatments[14]. In this review, we highlight the current insights of glioblastoma in the aspects of epidemiology, pathogenesis, etiology, molecular genetics, imaging technologies and treatment.

EPIDEMIOLOGY

The incidence of GBM slightly differs among the countries. According to the Australian Institute of Health and Welfare Data in 2013, GBM was the most common CNS tumors (62%). The new cases for 2017 in Australia were estimated as 2076 cases with a rate of 7.6 per 100000[15]. In United States, the Central Brain Tumour Registry of the United States report for the period of 2011 to 2015 showed that 57805 new cases of GBM were declared with a rate of 3.21 per 100000[16]. Xu et al[17] found a geographic variation among the United States regions with a higher incidence rate in the south region compared to other regions (per 100000/year: 24.31, 22.6, 20.35, 15.03 respectively in the South, Northeast, West, North Central regions). In the same study, it was reported that ethnicity has different susceptibility in risk of mortality with GBM, with the Caucasian being the most impacted, followed in this order by Hispanic, Asian and African. Some factors related to regional difference were the access to health care and rural vs urban areas. The epidemiological studies on Indian patients for all gliomas found that the incidence rate of tumor in India is between 1 to 4 per 100000[18]. A tertiary hospital in India found that 60% of all primary CNS tumors are a high-grade tumor[19]. GBM incidence in Brazil was accounted for 11% of all CNS tumors[20]. In Japan, the rate for gliomas occurrence was 6.6 per 100000 persons[21]. Astrocytic tumors of CNS in Europe had a rate of 4.4 per 100000 in the years between 1995 and 2002. It also found that GBM was less frequent in eastern Europe (38%)[22].

Survival rate usually documented for 1 or 5 years. The 5-year relative survival GBM was 4.6% in Australia, 4.9% in Europe, and 5.6% in the United States[16,22]. The 5-year survival rate varied by age and gender. Males have lower 5-year survival rate than females. A study in China on 6586 patients showed 6.8% and 8.3% 5-year survival for males and females, respectively[23]. Adolescents and young adults have a higher survival rate than older adults. Most of the studies showed that the frontal lobe is the most common site for GBM tumors[16]. The reason for this pattern of frequency may refer to the occurrence of tumors in larger lobes[24]. Despite the rapid developments of the standard treatment in clinical practice for GBM, the trend of survival rate did not change among years[12].

GENETICS AND PATHOGENESIS

Glioma is a term used to describe a group of cancers arising from glial cells of the CNS. Moreover, Glioma tumors are usually classified as infiltrative diffuse astrocytomas. GBM originally is derived from the star-shaped cells astrocytes and it is the most common diffuse astrocytomas[5]. The main characteristic features of GBM include tissue heterogeneity with multifocal hemorrhage, increased cellularity, nuclear atypia, angiogenesis, and mitotic activity. Necrosis, cystic and gelatinous regions were also observed within GBM[9,25]. A recent study showed that the neural stem cells within the subventricular zone of the brain are the cells of origin for GBM rather than mature glia[26]. Histologically, two types of gliomas were identified as the cradle of GBM: Astrocytomas and oligodendrogliomas[9]. GBM tumor cells have the capability to move a long distance from the tumor origin and stimulate additional tumor in another location[27]. Therefore, complete resection cannot be achieved, causing a tumor recurrence. Clinically, GBM can be divided into two types based on their origin: Primary and secondary GBM. Primary GBM arises de novo without any evidence of predecessor lesion. Secondary GBM arises from lower-grade tumors[9]. Some studies found that necrosis characterizes GBM in over 85% of the cases[28].

The genetic abnormalities associated with GBM were documented to distinguish between the primary and secondary glioblastomas[27]. Earlier molecular studies have identified the core pathways of GBM deregulation including receptor tyrosine kinase signaling, and the TP53 and RB tumor suppressor pathways[29]. The genetic alterations of primary GBM include epidermal growth factor receptor (EGFR) gene mutation and amplification, loss of heterozygosity (LOH) of chromosome 10q containing phosphatase and tensin homolog (PTEN), deletion of p16, and overexpression of mouse double minute 2. On the other hand, the genetic alterations of secondary GBM include platelet-derived growth factor receptor alpha (PDGFA/PDGFRa), overexpression of platelet-derived growth factor A, mutations of TP53 and RB, mutations of IDH1/2 and ATRX, and LOH of 19q[9,30,31]. The Cancer Genome Atlas Research provides more details about the genetics and molecular pathogenesis of GBM by identifying four subtypes: Classical, mesenchymal, proneural, and neural. The proneural subtype is characterized by CDK4, CDK6, PDGFRA, MET, and IDH1 mutations. The classical subtype is characterized by the loss of PTEN and CDKN2A and EGFR amplification. The mesenchymal subtype is characterized by mutations and/or loss of TP53, NF1, and CDKN2A. The last subtype neural does not have a unique genetic signature[9].

As stated earlier, the IDH gene, which makes an enzyme called IDH 1, represents the early genetic alteration in gliomas[32]. These genes are mainly found in cytoplasm and peroxisomes. IDH1-R132H mutation detection by immunohistochemistry was used to give the necessary molecular information of gliomas[33]. In a recent study, IDH1 mutation enhances the DNA damage response, which results in the radioresistance of the tumor[34]. Gene and nucleic acid carriers to penetrate GBM are under intensive studies in order to use gene therapy against GBM[35].

An essential biological characteristic of GBM is hypoxia. The principal gene transcription for hypoxia is HIFα. HIFα controls many genes, such as HRE-related gene transcriptions, EGFR, and others. Toyonaga et al[36] found that HIFα, VEGF, Ang2, and CAIX were higher in hypoxic gliomas.

In conclusion, further studies and research in genetics would provide more explanation for the diversity and heterogeneity of GBM.

ETIOLOGY

The etiology of GBM has not been entirely clarified. No risk factors can be clearly identified for GBM. Only 1% of the cases were reported to be linked to hereditary diseases. Multiple genetic diseases have been associated with GBM such as tuberous sclerosis, Turcot syndrome, multiple endocrine neoplasia type IIA, neurofibromatosis type I, and NF1[13,37,38]. No definite conclusions of the association of GBM with environmental factors such as smoking, dietary, head injuries, obesity, and electromagnetic field[9,39-41]. Exposure to ionizing radiation is one of the confirmed risk factors for GBM. High dose radiation was shown to increase the risk of glioma tumors while no association has been found with low radiation at the diagnostic level[40,41]. Some cases have been linked to the consequences of atomic bombs[42]. The effect of radiation-induced GBM was typically seen years after radiation therapy[43]. Contradictory results have been found with some chemicals such as pesticides, organochlorides, alkylureas combined with copper sulfates, vinyl chloride, petroleum refining, and synthetic rubber manufacturing[44]. Few studies suggested the relation of ovarian steroid hormones with GBM[45].

IMAGING MODALITIES

GBM imaging becomes an essential component in tumor grading and detection, treatment planning, treatment response, and patient prognosis. The functional imaging can help in providing additional information such as metabolism, perfusion, tumor proliferation and hypoxia[46]. Tumor treatment response usually assessed by using a response evaluation criteria system that is mainly based on changes in tumor volume. However, this system does not take the effect of functional and metabolic changes which may be evaluated using additional imaging modalities[47]. Treatment response could also be evaluated using tumor texture analysis which can evaluate the tumor heterogeneity[48]. Tumor angiogenesis was assessed mainly using functional nuclear medicine modalities such as PET and SPECT. These modalities were correlated with the most potent factor for angiogenesis VEGF[49]. Rainer et al[50] found that 123I-VEGF PET imaging can be useful for tumor angiogenesis evaluation in GBM patients.

MRI is the gold standard in detecting GBM tumors due to its superior soft tissue contrast. Many MRI techniques used for GBM including T1-weighted image, T2-weighted image, T1-weighted contrast-enhanced (T1CE), T2-fluid-attenuated inversion recovery (T2-FLAIR), diffusion tensor imaging (DTI), and gradient echo[51-53]. The MRI-T1 imaging with gadolinium enhancement can detect the malignant glioma with the necrosis (Figure 1). In most cases, the neurosurgeons would use high-resolution MRI for intraoperative guidance for an improved soft tissue contrast[54]. T2-FLAIR can detect 90% of GBM recurrence[55]. T1CE MRI becomes a routine after the surgery which helps to improve prognostic significance[56]. Besides, T1CE can measure the pharmacokinetic parameters of contrast uptake which is associated with early tumor progression[57]. Some recent studies with functional MRI showed the importance in surgical resection of GBM tumor which may disrupt eloquent areas[58]. DTI MRI may help in neurosurgical planning and can distinguish between vascular damage and residual enhancing tumor[59]. In radiation therapy treatment planning, MRI in registration with CT was used to define the tumor volume and organs at risk[60]. A recent study used MRI for the determination of IDH1 mutation status (mutant vs wild type) with the use of essential patient information such as age and tumor volume[61]. Another study was able to use MRI to predict the IDH1 mutations in GBM patients[62]. Spectroscopy techniques associated with MRI have been used to identify the extension of the tumor and areas of high risk of recurrence[63]. Response assessment using MRI is based on Macdonald criteria which classify the response into four categories (complete response, partial response, stable disease, and progressive disease) (Table 1)[64]. Many studies have reported the challenges and limitation of these criteria, mainly Sorensen et al[65] and Henson et al[66] studies highlight that the Macdonald criteria inadequately assess gliomas by focusing on contrast-enhancing areas, neglecting non-enhancing tumor components. Macdonald participated in another group who developed the Response Assessment in Neuro-Oncology (RANO) criteria, which has gained broader acceptance[67].

Figure 1 Example of Tl-weighted magnetic resonance imaging image and 18F-fluoromisonidazole positron emission tomography image. Both images show the tumor and the necrosis. A: Magnetic resonance imaging image; B: 18F-fluoromisonidazole positron emission tomography image.

Table 1 Macdonald criteria for treatment response based on major changes in tumor size on computed tomography or magnetic resonance imaging images[66].

Response	Criteria
Complete response	Complete disappearance of disease and no new lesions
Partial response	≥ 50% decrease compared with baseline volume and no new lesions
Stable disease	Does not qualify for complete response, partial response, or progression; clinically stable
Progression	≥ 25% increase of the baseline volume, any new lesion, or clinical deterioration

CT is often used for visualizing the tumor. CT is usually the first choice for some patients who cannot have MRI such as patients with pacemakers. GBM tumors with CT appear with less intensity compared to surrounding tissues[13]. A recent study showed that spectral CT can detect small changes with GBM and provides relevant information regarding the proliferation and invasion of cancer cells[68]. However, CT remains the mostly used imaging modality in cancer treatment planning.

PET is a technique in nuclear medicine imaging which uses specific radiopharmaceuticals for the diagnosis and monitoring of tumors. PET provides additional features which cannot be imaged with MRI or CT. Some of these features are tumor grading, delineation, differential diagnosis, surgery planning, and radiotherapy planning[46]. Many radiopharmaceuticals have been studied with GBM (Table 2). 18F-2-fluoro2-deoxy-D-glucose (18F-FDG) is widespread used with GBM. It was used to study the differential of low-grade glioma, identify anaplastic transformation, and reveals information regarding the prognostic value[69,70]. However, 18F-FDG has high uptake in normal brain tissue which reduces tumor contrast and delimitation[71]. Amino acid tracers can detect gliomas with higher sensitivity than 18F-FDG[72]. They also help in the differentiation of recurrent tumors from treatment-induced tumors, and they have shown higher sensitivity at the initial stage of GBM. One of the most important amino acids studied with GBM is 11C-methyl-methionine (11C-MET). The main drawback with 11C-MET is the short half-life requiring a nearby cyclotron. Some labeled 18F amino acid have been developed to overcome the shortage of 11C such as O-[2-(18F)-fluoroethyl]-L-tyrosine (18F-FET) and 3,4-dihydroxy6-(18F)-fluoro-L-phenylalanine (18F-FDOPA)[73]. 18F-FET and 18F-FDOPA have shown similar results to 11C-MET in brain gliomas detection. Tumor delineation of GBM has shown greater possibility with amino acid PET than with 18F-FDG[74]. A study showed that 18F-FET could detect the tumor mass and reveal residual tumor more reliably than MRI[75]. 18F-FDOPA PET has shown similar results compared with MRI[76].

Table 2 Positron emission tomography radiopharmaceuticals used with glioblastoma multiforme.

Biological processes	Radiotracer
Primary diagnosis-differential diagnosis–treatment planning and response	18F-fluorodeoxyglucose (18F-FDG)
	11C-methyl-methionine (11C-MET)
	O-(2-18F-fluoroethyl)-L-tyrosine (18F-FET)
	3,4-dihydroxy6-18F-fluoro-L-phenylalanine (18F-FDOPA)
	18F-fluorotyrosine (18F-TYR)
	18F-fluoromethyltyrosine (18F-FMT)
	18F-fluorothymidine (18F-FLT)
Tumor hypoxia	18F-fluoromisonidazole (18F-FMISO)
	18F-fluoroazomycin-arabinozide (18F-FAZA)
	18F-fluoroerythronitroimidazole (18F-FETNIM)
	18F-2-nitroimidazol-pentafluoropropyl acetamide (18F-EF5)
	18F-2-nitroimidazol-trifluoropropyl acetamide (18F-EF3)
	18F-fluoroetanidazole (18F-FETA)
	124I-iodoazomycin galactopyranoside (124I-IAZGP)
	68Ga-labeled nitroimidazole analogs
	60,61,62,64Cu-diacetyl-bis (N4-methylthiosemicarbazone (60,61,62,64Cu-ATSM)
	124I-chimeric mAb G250 (124I-cG250)
	89Zr-chimeric G250 F(ab′)2 (89Zr-cG250-F(ab′)2)
Perfusion-membrane biosynthesis-proliferation rate	13N-NH3
	18F-fluorthymidine (18F-FLT)
	18F-fluorocholine (18F-FCH)
	15O-labeled water (15O-H2O)

Other tracers have been developed for GBM based on physiological intention. For instance, the determination of hypoxic tumor volume can help in improving the overall treatment response. PET detection of hypoxia was considered the preferred technique due to its high sensitivity and specificity[77]. For hypoxia, 18F-FDG was used initially to measure the hypoxic volume[78]. Nitroimidazole compounds are the most studied for hypoxia measurement. The mostly used tracer was 18F-fluoromisonidazole (18F-FMISO) (Figure 1). 18F-FMISO was shown to accumulate in high hypoxic tissues[79]. FMISO has also been used to predict the tumor necrosis in GBM[80]. Besides, hypoxic glucose metabolism has been found to be a significant predictor for GBM tumors[80]. The low tumor to background ratio and the slow kinetics considered the significant drawbacks of 18F-FMISO. 18F-fluoroazomycin-arabinozide has faster kinetics compared to 18F-FMISO in detecting tumor hypoxia but has limited evidence[81]. Other tracers for hypoxia are presented in Table 2. Another radiotracer, ammonia 13N-NH3 has been used in combination with 18F-FDG for GBM perfusion and metabolism[82].

Sometimes, the analysis of PET images requires additional quantification techniques relating the activity concertation with PET to specific physiological behavior. The quantification techniques based on the acquisition classified into two modes: Static and dynamic. Static acquisition modes are measured at fixed sample time. In clinical practice where image acquisition is in static mode, the most widely used method for static quantification is the Standard Uptake Value (SUV)[83]. SUV is a semi-quantitative ratio defined as the ratio of activity in a region of interest (ROI) to the whole injected activity divided by patient weight or patient body surface area. It can distinguish between normal and abnormal uptakes, and it is simple to apply. However, studies have shown that SUV is affected by many technical and physiological factors such as the time of the measurement, the accuracy of the injected radiotracer activity[84]. Another static method used is the reference ratio method. Reference ratio method is the ratio of activity in an ROI to the activity in a reference tissue. Normal tissue and blood volume were used as reference tissue in PET[85–87]. Similar to SUV, the ratio methods do not provide information about the kinetics of the tracers in GBM tumors.

For the dynamic method, the mostly used approach for GBM is compartmental modeling (CM)[88]. This method allows to estimate the micro and macro parameters related to tracer’s physiology. Another alternative to CM is the graphical analysis methods of dynamic data such as and Logan et al[89] and Patlak and Blasberg[90] analysis. Their popularity came from their simplicity and model independence. Nevertheless, graphical methods estimate only a single parameter (net trapping uptake or distribution volume). Also, some studies reported the effect of noise on graphical methods. Another method for analysis of dynamic PET images is the spectral analysis (SA) method[91,92]. SA provides a spectrum of the estimated components from which physiological parameters can be derived. The SA does not require any prior information about the system, and it can help in obtaining the number of components in each voxel. However, the image components are usually corrupted by noise from the whole image (Figure 2).

Figure 2 Tumor-to-blood ratio, tumor-to-reference tissue ratio and spectral analysis images for two subjects S1 and S2 scanned with and 18F-fluoromisonidazole at 3 hours post-injection. Since tumor-to-blood ratio depends on the precision of the blood function, hypoxia delimitation was not always clearly defined in comparison to tumor-to-reference tissue ratio which was able to isolate the high intensity tumor pixels from the background. Spectral analysis still depends on the selected exponentials forming the tumor including noisy pixels[87]. TBR: Tumor-to-blood ratio; TNR: Tumor-to-reference tissue ratio; SA: Spectral analysis.

Imaging of GBM remains the most convincing approach regardless of its availability and expense. Imaging with PET in dynamic mode provides multifaceted ways of analyses. The ultimate goal is to delimit hypoxic regions of a tumor[85,86]. Figure 2 shows two GBM tumors from two patients where the hypoxia delimitation was attempted by means of tumor-to-blood ratio, tumor-to-reference tissue ratio and by means of SA. These images were acquired with 18F-FMISO at 3 h after radiotracer injection.

TREATMENT

Treatment of GBM tumors is still a challenging issue in oncology[93]. Despite the new developments in cancer treatment, the survival rate did not significantly improve over the last two decades[12]. The current standard techniques for the newly diagnosed GBM include maximal surgical resection, followed by radiation therapy and chemotherapy[13]. The management of GBM patients does not only include therapeutic management but also the supportive care for patients with neurological symptoms and other disease symptoms[94]. The treatment options depend on several factors related mainly to tumor size, time of diagnosing, and patient age[95]. The RANO criteria was used to evaluate the treatment response[96]. Usually, MRI is the main modality for the follow up of GBM treatment response. Advanced techniques of MRI can monitor radiation necrosis, pseudo-progression, and tumor progression[97]. PET imaging has potential applications in the monitoring of treatment response. 18F-FET PET can differentiate tumor recurrence from radiation necrosis[98]. Besides, multiple hypoxia imaging during chemoradiotherapy treatment course can demonstrate the extent of hypoxia and may help in modifying the patient treatment plan[99-101].

Surgery can help in reducing the tumor burden, control symptoms, and application of in vivo therapeutic agents. Due to the multifocal pattern of GBM in the CNS, a complete and extensive surgical resection of the whole tumor is difficult to achieve. Therefore, the removal of the primary tumor is not curative, and some tumor cells remain within the brain which could lead to tumor progression or recurrence[102,103]. Some regions in the brain cannot be removed by surgery such as tissues in eloquent cortex, brain stem or basal ganglia.

Following the surgery, radiation therapy can be applied after the craniotomy wound has healed. It has been shown that radiation therapy improved the survival rate of GBM patients[104]. Multiple technologies were used in radiation therapy such as brachytherapy, stereotactic radiosurgery, intensity-modulated radiation therapy (IMRT) and boron neutron capture therapy (BNCT)[94,105]. Treatment with IMRT and BNCT have been shown to reduce the dose to healthy tissue within the brain. The standard total dose delivered to GBM with IMRT is around 60 Gy, a 2 Gy-fraction per day over six weeks[94].

Recent studies on tumor resection surgery have explored the use of hydrogel implants in the resected tumor cavity as a strategy to either attract tumor cells from surrounding healthy tissues or deliver targeted therapies directly to the area[106,107]. In certain treatment protocols, the hydrogel—sometimes incorporating nanoparticles—is subjected to irradiation. This approach helps to reduce the risk of tumor recurrence by preventing the regrowth of residual cancer cells, thus improving the effectiveness of the surgery and enhancing local tumor control. Additionally, the combination of hydrogel and irradiation can potentially serve as a platform for sustained drug release, promoting continuous therapeutic action while minimizing systemic side effects.

Temozolomide (TMZ), a chemotherapeutic agent to treat GBM, was a standard since 2005. After that, the use of radiation and TMZ became the standard of care for treating GBM patients[108]. The survival rate improved from 12.1 for radiation alone to 14.6 for radiation with TMZ[104]. Intra-arterial chemotherapy, a novel technique to increase the local concentration of a drug with unfortunately high toxicity to healthy tissue, has been used for GBM. The overall median survival rate was 25 months[109].

The TMZ is given orally at a dose of 75 mg/of body surface area daily for six weeks with radiation. After a rest period of one month, a restarted dose of TMZ with around 200 mg is given for five days monthly. The total time of TMZ administration range from 6 to 18 months[110]. Despite the approval of using TMZ for GBM patients, several side effects have been documented for TMZ. Table 3 Lists some common chemotherapeutic agents used for GBM.

Table 3 Chemotherapeutic agents commonly used for glioblastoma multiforme.

Drug	Dose	Ref.
Temozolomide	75 mg/m2 daily and 150-200 mg/m2 (5/28 days)	[106]
Bevacizumab	10 mg/kg every 2 weeks	[110]
Procarbazine	110 mg/m2 day 1/56	[113]
Vincristine	1.4 mg/m2 days 8 and 29/56	[113]
Lomustine (CCNU)	60 mg/m2 days 8-21/56	[114]

Tumor treating fields (TTF) is a new technique for recurrent GBM treatment having the approval of the United States Food and Drug Administration in 2011. This device emits low-intensity, intermediate-frequency (100–300 kHz) alternating electric fields causing mitotic cell death. The median survival with TTF was 6.6 months vs 6 months for TMZ only[111].

Cisplatin is a type of platinum chemotherapeutic drug used in cancer treatment[112]. It has been cited as being the most cytotoxic drug because of its various applications in the treatment of many types of tumors[2]. For brain cancer, Cisplatin was used for recurrent childhood brain tumors[113]. It is the standard agent for medulloblastoma but not for glioblastoma tumors[114]. There is no clear explanation about the differences of Cisplatin efficiency in medulloblastoma and glioblastoma, with conflicting results in the effect of intratumoral administration[115]. Current studies compare the effectiveness of adding Cisplatin with TMZ. Seidel et al[116] suggested the use of TMZ only for glioblastoma due to the high toxicity of Cisplatin. However, a novel study used nanocarrier for Cisplatin delivery found negligible toxicity[117].

Cetuximab is another chemotherapeutic drug used in cancer treatment. It is an antibody that binds to the EGFR, which is expressed in many types of cancer. It has been approved for colorectal cancer and squamous cell carcinoma of the head and neck[118]. Animal studies found promising the effects of Cetuximab on glioblastoma models[119]. However, a phase II study in 2006 failed to demonstrate the efficiency of Cetuximab in combination with TMZ and radiation in the treatment of the primary glioblastoma[120]. An ongoing phase II research studies the effect of intra-arterial infusion of Cetuximab for the treatment of newly diagnosed glioblastoma with no results published till now[121].

Other technologies have been investigated such as radioimmunotherapy, iodine-125 brachytherapy, and hyperfractionation, however, the results did not significantly improve the survival rate[43].

Several new avenues to treat GBM were suggested in the literature among them the use of nanoparticles and antibodies. Nanoparticles can be made of small vesicles about a nanometer in size and encapsulating the drugs to be delivered to the target tumor and reducing side effects[122]. The nanoparticles can be delivered to the tumor through its porous vasculature by passive diffusion, and they accumulate in the tumor due to its inefficient clearance, a concept known as enhanced permeability and retention effect. The release of the drug is controlled by the engineering process of the nanoparticles, and the drug can be delivered internally to the cells[122,123]. Concomitantly, the nanoparticles can be designed to provide images of their distribution in the body.

Antibodies are proteins produced by the immune blood plasma cells to act against pathogens like bacteria and viruses by recognizing their antigen. Monoclonal antibodies used against cancer are engineered to target specific antigens in cancer cells. They directly act on cancer cells, or they affect tumor blood vessels. The monoclonal antibodies can be carrier of radioactive substances or drugs. These are called conjugated monoclonal antibodies. By locally and selectively delivering the drugs to the cells of cancer, monoclonal antibodies are supposed to be more efficient than chemotherapy, while when transporting radioactive substances to target cancer cells, this approach is less confined to the tumor because the emitted radiation by the isotopes can interact farther than the tumor and might affect the normal surrounding tissues. In addition to delivering the radiation to the tumor, the same radiation emitters can be used for imaging. This is the case where monoclonal antibodies carrying and emitters like zirconium-89 and copper-64 are used for both treatment and imaging. Such combined treatment and imaging for diagnostic is called theranostic. Copper-64 has a half-life of 12.7 hours and emits at an energy of 0.579 MeV with a branching fraction of 38.4% and at an energy of 0.653 MeV with branching fraction of 17.8%. Although the monoclonal antibody is specific to the type of cancer, copper-64 can provoke irradiation to cells beyond the tumor by both and the imaging necessitates a high activity in order to provide images with acceptable contrast, which could be, at the end, harmful to the patient.

One main problem facing such therapies, nanoparticles and antibodies, is the transport of the drug through the blood-brain barrier (BBB). The BBB selectively protects the brain at the capillary endothelial membrane from toxins and other undesirable molecules harmful to the brain. To overcome the BBB and deliver the drugs to the brain, the assembling structure of the drugs are modified while keeping intact their active efficiency, the drugs can be delivered by neurosurgery, or by opening the BBB which can be mainly achieved by means of osmotic pressure, and microbubbles combined with ultrasound. To cross the BBB a drug molecule should be less than 180 Da and be sufficiently liposoluble[124]. Opening the BBB by osmotic pressure allows to open the whole brain hemisphere and this method is already used in the clinic. This method uses molecules such as arabinose, hyperosmolar solution of mannitol and lactamide. The action is made on the rearrangement of the endothelial cells of the capillaries[124]. Focused ultrasound in conjunction with injected microbubbles in the vascular space allow to open the BBB for the whole brain[125]. This approach is under active research in large animals.

Among the recent advances in cancer treatment is the development of therapies that inhibit DNA repair mechanisms, which can help to target cancer cells more effectively by preventing their ability to fix genetic damage. The target of this approach is the Poly (adenosine diphosphate ribose) polymerase (PARP) family, which is involved in the repair of DNA strand breaks caused by radiation therapy within the tumor. By inhibiting the action of PARP inhibitors, tumor cell's ability to repair DNA damage is disrupted, provoking cell's death or inability to proliferate. Drugs such as Niraparib, Rucaparib, and Olaparib are all PARP inhibitors, a class of drugs used in cancer treatment, particularly for cancers with defects in DNA repair mechanisms like BRCA1/2 mutations as in ovarian, breast, pancreatic, and prostate cancers. The combination of PAPRi inhibitors like Olaparib and TMZ was reported to be more efficient in tumor cell killing compared to either drug alone[126,127].

HEALTHCARE COSTS OF PATIENTS WITH GBM

According to a United States study, a total of 2272 patients received cancer treatment between 2006 and 2010, the total healthcare costs were between $79099 to $138767 per patient. This range depends on the type of treatment received by the patients. 37% of patients received chemotherapy with radiation, 13% received only radiation, 3.9% received chemotherapy alone, and the remaining had only surgery[128]. Another United States study in 2921 patients found the estimated cost to be around $268031. This study adds the increasing healthcare cost until five years[129]. A recent study found a similar healthcare cost of about $185000 with a high contribution referred to radiation therapy and longer survival rate for patients with commercial insurance[130].

THE EARLY DETECTION IS THE KEY

The fight against cancer has been conducted through molecular and genetic approaches, through surgery, chemotherapy and radiotherapy, and even with environmental behavior like diets, physical activity, and cessation of tobacco and alcohol. It is well known that GBM develop discretely, and more importantly, GBM develop heterogeneously[131], and when the symptoms appear, the cancer is already spread. However, GBM do not generally metastasize to other tissues outside the brain because of the blood brain barrier[93]. When GBM, in all patients, form hypoxia, necrosis, angiogenesis and other characteristics, is this a sufficient justification to equally treat all patients with 70 Gy in radiotherapy? A personalized treatment based on early detection of the disease by means of family history, lifestyle, circulating tumor cells or other biomarkers[132] with the contribution of artificial intelligence (AI) deep learning (DL) encompasses all neural network-based methods, including transformers), and the potentially usage of vaccines in the very near future[133], conjugated with actual detection approaches and treatments, will be of great hope.

HOW AI CAN CONTRIBUTE TO DEFEATING BRAIN CANCER

Cancer cells can dynamically switch between different phenotypic states or cellular identities, depending on factors like genetic expression, microenvironment, and epigenetic changes. They adjust their spatial cellular organization, resist therapies and drive tumor progression. In order to characterize a tumor with DL, several thousands of samples have to be previously analyzed from which DL is trained. For example, in a study on GBM, Zheng et al[134] used 40 million tissue samples from 410 patients to associate tumor histologic architecture with prognosis using DL. In such a case, several research groups or hospitals have to make available their confirmed annotated data by experienced clinicians, and at the same time, they have to protect the confidentiality of the patients. Anonymization involves more than just removing text data from the image file header or other media; it also includes ensuring that the image itself does not allow for the identification of a patient. With such diversity of the data, DL has the potential to segment the tumor images, to classify and characterize the tumors, to perform tumor grading, and to provide prognosis prediction including response to treatment and survival estimate[135].

The main challenge remains that GBM tumors are often well advanced by the time they are diagnosed. In this regard, there exist a few statistical and AI models to predict the development of GBM based on factors like genetics, family history, age, lifestyle and environmental exposures, but the results are still not satisfactory. GBM is influenced by a complex set of factors that are not yet fully understood, and the influence of the environmental conditions are also difficult to assess.

Although the prediction of GBM development remains difficult, the planification of the treatment is possible with DL. Personalized factors in addition to medical images and biological samples will be in near future exploited by DL to individually propose appropriate treatment plans to a single patient. Depending on each patient, GBM tumors' shape and structure, primarily hypoxia, are different[85] and respond differently to therapies, while their typical irradiation dose in radiotherapy treatment is around 60 Gy administered in several fractions concomitant with chemotherapy and, in certain cases, followed by adjuvant treatments[101]. DL can detect structures within the tumor not discernible to radio-oncologists. It can segment a tumor based on its perfused, or hypoxic spots, and suggests appropriate radiation doses to different structures, thus helping medical physicists in the dose adjustment (dose painting).

Radiomics was limited to imaging like CT, MRI, PET and X-rays, where it involves collecting a variety of features from images, such as texture, shape, intensity, and other statistical data. Nowadays, radiomics has expanded to include other types of data such as genomics, pathology molecular characteristics, and clinical data, to develop personalized treatments and improve outcome prediction. Gathering information through different strategies, and from multinational cohorts, is very costly in terms of both subsidies and time, and some approaches are invasive like histopathology and tumor molecular profiling, which are even limited due to tumor heterogeneity. However, such data fed to DL makes a great difference when compared to radiomics. While DL doesn't currently predict the occurrence of GBM, it certainly improves its diagnosis, personalized treatment planning, outcomes and patient follow-up which results in an improved survival rate and a better quality of life.

Without delving deeply into the technical details of AI architectures, these techniques have been demonstrated to operate with fine details and speed that surpass human clinical capabilities. Their potential appears unlimited. While they are already contributing to the imaging-based diagnosis of GBM, their future potential lies in enabling predictive analytics, and more importantly to provide knowledge at the cellular and molecular levels. Ultimately, AI will help in designing simulations of the disease and suggesting treatments, bypassing, for certain assessments, the use of laboratory animals. The simulations can be from an external level like generating images from real existing and annotated images with clinicians, or the images are totally invented. A more important simulation is the one based on genes, environment and behavior. This case of simulation starts in the cells, in the same way like growing a tumor from implanted cells in a small animal.

For GBM image segmentation, the appropriate images for this task are better obtained with MRI, as it detects the tumor structural and functional properties with higher contrast and spatial resolution in comparison to other imaging modalities. In this context, different MRI imaging sequences are accordingly planned to delimit the tumor or to assess its metabolism[136], namely, for its structure: T1-weighted with contrast agent to enhance tumor area and hypoxia; T2-weighted for edema, fluid and soft tissues; FLAIR for infiltrative tumor. And for its metabolism: Diffusion-weighted imaging to measure cellularity through water motion; Diffusion tensor imaging to map white matter tracts; Perfusion to assess blood flow and vascular permeability; MR spectroscopy to measure concentration of brain metabolites. These images inherently lead to different visual representations of GBM. In traditional image segmentation like thresholding, region growing, edge detection and watershed segmentation which can be operator supervised or unsupervised, these approaches are often limited or inaccurate in glioblastoma due to intensity inhomogeneity, noise, and variability across patients. Combining two or more MRI imaging sequences as inputs to AI, the tumor can be more accurately delimited even after resection and after daily radiotherapy[137].

Moreover, an image of a brain with GBM can be decomposed into two or more images depicting different states of the tumor like necrosis, hypoxia and perfused regions, in similar fashion to dynamic PET image decomposition with factor analysis, SA and independent component analysis[85,138,139]. Image decomposition can be seen as a partition of image voxels into their constituents. In fact, when applying three-compartment pharmacokinetic modeling to PET dynamic images at the voxel basis, in the example of 18F-FDG, the results are provided in several images representing several parameters: Perfusion, a single or a combination of rates constants, and any of the three compartments. Since AI is actually able to generate synthetic images, no doubt it has the potential to generate images with higher spatial and temporal resolution than provided by the imaging scanners. Basically, when considering image formation with PET or MRI, the radiotracer or radiofrequency action in blood, tissues and cells is almost known, and can be simulated with AI, to finally produce similar images as those generated by the scanners. The goal is not to produce similar images with AI, but by reproducing scanners' images by calculations at the cellular and molecular levels, thereby matching scanners' precision and validating its accuracy, AI is then used to report both detailed and isolated cellular-scale interactions that the scanners cannot directly measure, and this is a way of generating more resolved and decomposed images at the same time.

Identical approaches can be applied to treatments by radiotherapy and chemotherapy. In radiotherapy, Monte Carlo simulations are well established methods in tumor targeting and for dose calculation. They are even used to depict the radiation interactions at the atomic level[140]. Such calculations can be used to train AI for more accurate and faster absorbed dose calculation, and for improved normal tissue sparing.

Chemotherapy methods offer much to discuss. They extend from drug discovery and development, personalized treatment such as genomic sequencing data to suggest targeted therapies like IDH1 gene mutation in GBM, to adverse effects of cancer drugs in specific patients. These methods can be illustrated by a work reported by Tran et al[141] where they described a computational drug repurposing strategy. The hypothesis was that if a drug causes gene expression changes similar to those seen with known anticancer drugs, it might also have anticancer effects. Among 1815 approved drugs obtained from the LINCS consortium having 166164 drug response data, they found one of them (Z29077885) which was not previously tested as anticancer reagent. They subsequently tested this compound and demonstrated its anticancer potential targeting STK33. Other approaches include screening with AI millions of compounds to find promising candidates, predict drug-protein interactions, and design new drug based on the targets' nature.

CONCLUSION

Even though the current treatment options for patients with GBM are still not satisfactory, there are hopes of some new developments of cancer treatments to increase the survival rate. Besides, the latest techniques in medical imaging used with GBM may improve the diagnosis, tumor delineation, and treatment planning decisions. Apart from diagnosis and treatment approaches, a method to early detect GBM would drastically improve the overcome of the battle against this disease. AI can fulfill such a role by relying on genetic predisposition, family history, lifestyle and biomarkers suggesting early screening for potential high-risk individuals.