MCE: Medical Cognition Embedded in 3D MRI feature extraction for advancing glioma staging

Introduction Brain tumors pose a significant threat to global health and quality of life. Among them, gliomas are the most common type of malignant brain tumor [1]. Grading of gliomas, a crucial factor in clinical diagnosis and treatment, is primarily based on histopathological analysis, which can be supplemented by imaging evaluation using Magnetic Resonance Imaging (MRI) [2]. MRI can clearly demonstrate the anatomic structures of the brain, thereby enabling the precise localization and delineation of tumors. It plays a pivotal role in the detection, classification, and staging of brain tumors [5]. However, the current means of assessing and staging brain tumors through MRI images rely heavily on experienced radiologists, which can be susceptible to subjective bias and human errors, especially when dealing with large volumes of image data [7, 8]. Recent advances in deep learning algorithms have achieved revolutionary breakthroughs in the automated analysis of brain glioma MRI images, providing new opportunities for accurate detection and classification [9]. These algorithms can learn complex features from MRI data and perform well on classification and staging tasks [10-12]. However, limitations also exist in purely data-driven deep learning algorithms, including (1) the lack of guidance from prior knowledge and domain-specific rules in MRI analysis, resulting in excessive redundancy in automatically learned features and patterns, which in turn affects the model's generalization ability [13]; (2) while deep learning features can effectively capture high-level abstract information in images, they have limitations in expressing specific features relevant to the staging of gliomas, such as texture, morphology, and grayscale distribution; (3) models may mistakenly identify irrelevant image features, such as artifacts, vignetting, background, and the like, as decision cues, indicating the model's excessive reliance on the specificity of training data [14].

To address these limitations, we propose a new method that incorporates external knowledge rules from the perspective of feature acquisition. Specifically, we embed knowledge characteristics into data-driven approaches to enhance the quality of feature extraction. Given the limitation of 2D-based analysis methods, our approach is tailored for 3D brain glioma MRI images. Our research introduces a novel brain glioma staging model called the Medical Cognition Embedded (MCE) model for 3D data. This model uses a combination of different modules to improve the generalization capability and enhance the quality of feature extraction. These modules include: (1) Multi-Sequence Attention Neural Network (MSANN), which imitates the attention pattern of physicians during the interpretation of medical images, focusing on the characteristic regions of different sequence images; (2) a Clinical Learning Neural Network (CLNN) is designed to mimic the learning process of a physician during film reading; (3) a Radiomics module that extracts and calculates specific features from different sequence images to enhance the predictive performance of the model; and (4) manually created diagnostic features (Diag-Features), which incorporate key diagnostic insights to further enhance the performance of the model. The efficacy of the proposed methodology is rigorously evaluated on the publicly available BraTS2018 and BraTS2020 datasets, comparing it to other well-known data-driven models and achieving higher accuracy, recall, and precision, reaching 96.14%, 93.4%, 97.06%, and 97.57%, 92.80%, 95.96%, respectively.

Related works Feature engineering is a key component of MRI analysis. Features used for brain glioma staging tasks include deep features and manual features. Deep features are featured extraction networks designed to extract features from images, thereby enhancing classification performance through improved classifiers or training strategies. These features can effectively capture both local and global information in the images [15]. However, manual features require experts to design feature calculation methods based on specific medical knowledge and task requirements, such as statistical, morphological, and textural aspects (e.g., LBP, GLCM, etc.) for manual extraction [20]. Each feature has its physical significance, thus manual features are often highly interpretable. These features are typically combined with machine learning algorithms such as Support Vector Machine (SVM), Random Forest, etc. [21].

To address the challenges of small sample glioma imaging data, the literature has primarily relied on techniques from computer vision, such as data augmentation, simplifying model complexity, expanding the sample space through generative techniques, and improving classifier algorithms [22]. However, from the perspective of external information acquisition, these methods primarily focus on optimizing the target task within the given dataset and do not introduce new external information into the model.

To address these problems, embedding knowledge characteristics into data-driven approaches can help the model better understand and utilize small-sample data [23]. There are two broad categories: the first is low-level knowledge, which includes strategies such as image normalization, data augmentation, and transferring feature extraction models from large-scale natural image datasets or other medical images. The second direction is high-level knowledge, which includes strategies such as domain-specific rule transfer and sample enrichment through knowledge graphs or word embeddings [24].

Our work primarily focuses on the second direction and transfers knowledge characteristics into data-driven approaches to improve the quality of feature extraction in limited datasets. We achieve this by designing the MCE model, which incorporates multiple modules to enhance the generalization capability and the quality of feature extraction. These modules include MSANN, CLNN, a Radiomics module, and manually constructed Diag-Features.

Materials and methods The first chapter introduces the background and innovative aspects of our work. The second chapter presents relevant works on glioma staging and the concept of knowledge embedding, while also discussing the role played by knowledge rules in feature extraction. The third chapter introduces the materials and methods used

Read more