The Role of Machine Learning in Preoperative 3D Planning and Mapping for Brain Tumor Resection: A Review

Preoperative planning is a critical step in neurosurgical oncology to maximize brain tumor resection while preserving eloquent cortical areas. Traditional 2D magnetic resonance imaging (MRI) analysis is time-consuming and prone to human error. This review examines the current state of artificial intelligence (AI) and machine learning (ML) algorithms in automating tumor segmentation, 3D structural mapping, and surgical pathway planning. By analyzing recent clinical literature (2022–2026), we assess the accuracy, limitations, and future outlook of human-AI collaboration in neurosurgery. Surgical resection remains the cornerstone of treatment in neurosurgical oncology for both highly aggressive primary malignancies, such as high-grade gliomas [2, 10], and complex benign lesions like meningiomas [8, 9]. However, executing a safe and maximal tumor removal presents a formidable challenge due to the infiltrative nature of brain tumors, which often lack distinct macroscopic boundaries and intimately border critical eloquent cortical zones responsible for speech, motor functions, and cognition [4]. Traditional preoperative planning relies extensively on the manual interpretation of multi-sequence and functional magnetic resonance imaging (fMRI) datasets [1]. This clinical approach is not only highly time-consuming but also inherently subjective, introducing a critical reliance on human experience and increasing the margin for human error during boundary delineation [3, 6]. Limitations of Traditional Approaches and Critique of Manual Processing Surgical resection of brain tumors presents a critical clinical challenge due to the delicate balance between maximizing tumor removal and preserving vital brain functions. Historically, neurosurgeons relied on conventional two-dimensional (2D) magnetic resonance imaging (MRI) scans, attempting to mentally reconstruct a three-dimensional (3D) representation of the patient's brain anatomy. This manual process is time-consuming and inherently vulnerable to human factors, such as visual fatigue, which increases the risk of overlooking exact tumor boundaries. The limitations of manual evaluation become increasingly pronounced when handling complex advanced diagnostic modalities. Functional MRI (fMRI)-encompassing task-based fMRI (tb-fMRI) and resting-state fMRI (rs-fMRI)-is increasingly applied for noninvasive preoperative mapping of widely distributed brain areas. However, manually interpreting and synthesizing such vast and complex data streams presents major technical challenges: Functional signals (blood oxygen level-dependent / BOLD) in rs-fMRI are particularly susceptible to tumor-induced alterations, technical noise, and substantial data processing challenges, making purely manual interpretation unreliable without systematic processing. Comparative validation studies between fMRI and intraoperative direct cortical stimulation (DCS) demonstrate variable levels of concordance, particularly regarding cognitive functions such as language (with tb-fMRI generally exhibiting higher concordance than rs-fMRI). Relying solely on manual oversight without automated integration complicates the accurate synthesis of these divergent findings. Furthermore, while preoperative fMRI-when combined with MRI tractography and intraoperative DCS-holds potential to guide surgical planning, improve the extent of resection, and reduce functional deficits, prospective evidence proving its direct impact on survival and patient outcomes remains limited. Consequently, traditional manual methods lack the consistency required to efficiently process complex multi-modal diagnostic data, highlighting the necessity for standardized analysis techniques [1] To address the constraints of traditional manual data analysis, artificial intelligence (AI) and deep learning (DL) are increasingly being integrated into clinical practice to enhance medical management, particularly in personalized diagnostic and treatment strategies for brain tumors. The rapid accumulation of imaging records and medical history presents massive datasets that easily overwhelm manual methods. DL models effectively shorten diagnosis time by extracting relevant clinical information from these large volumes of data. Advanced deep learning architectures, including two-dimensional and three-dimensional (3D) convolutional neural networks (CNNs) as well as DenseNet, facilitate precise radiological diagnostics across modalities such as computed tomography (CT), positron emission tomography (PET), and magnetic resonance imaging (MRI). By automated tumor segmentation, classification, and precise 3D tumor delineation, these algorithms identify tumor phenotypes and brain metastases. Consequently, such computer-assisted interventions enable real-time decision-making, enhance preoperative planning, and provide accurate, repeatable image analysis necessary for guiding complex neuro-interventional procedures[3,10].