Deep learning outperforms classical machine learning methods in pediatric brain tumor classification through mass spectra

Thais Maria Santos Bezerra, Matheus Silva de Deus, Felipe Cavalaro, Denise Ribeiro, Ana Luíza Seidinger, Izilda Aparecida Cardinalli, Andréia M. Porcari, Luciano de Souza Queiróz, Hélio Pedrini, João Meidânis

Abstract

Pediatric brain tumors are the leading cause of cancer-related death in children, with surgical resection serving as the primary management strategy. Achieving safe gross resection—maximizing tumor removal while preserving healthy tissue—is critical, yet current intraoperative margin delineation methods lack the necessary speed and precision. This study evaluates the efficacy of machine learning and deep learning models in classifying pediatric brain tissue based on chemical profiles generated by Desorption Electrospray Ionization Mass Spectrometry Imaging (DESI-MSI). Analyzing 34,870 annotated spectra from 116 patients, we compared classical methods (SVM, Random Forest, LASSO) against custom deep neural networks. Our results demonstrate that deep learning significantly outperforms classical techniques, particularly in the challenging task of distinguishing between low-grade and high-grade malignancies, achieving 87.1% accuracy compared to the ~65% seen in traditional models.

Introduction

In pediatric neuro-oncology, the surgical objective is often "gross total resection." However, identifying the exact boundary between neoplastic and healthy tissue is a major clinical challenge. Traditional intraoperative pathology can be time-consuming and invasive. DESI-MSI offers a compelling alternative by providing a rapid chemical "fingerprint" of tissue based primarily on its lipid profile. Because tumor cells exhibit distinct metabolic phenotypes and "lipid avidity," these spectra contain high-dimensional data that can theoretically be used to automate tissue classification.

Mapping mass spectra to tissue microanatomical regions using DESI-MSI.

Methodology and Datasets

The study utilized a robust dataset from the Boldrini Children’s Center, comprising 34,870 annotated mass spectra collected from 116 pediatric patients over two decades. Each spectrum consists of over 12,000 features representing signal intensities across a mass-to-charge ($m/z$) range of 180–1200.

Data Stratification

To prevent "data leakage"—where a model might simply memorize a specific patient's profile rather than learning generalizable features—the researchers employed StratifiedGroupKFold cross-validation. This ensured that all spectra from a single patient remained together in either the training or testing set.

Distribution of annotated spectra across different tissue findings and malignancy grades.

Model Architectures

Two distinct approaches were compared:

  1. Classical Machine Learning: Support Vector Machines (SVM), Random Forests (RF), and LASSO regression were evaluated alongside a standard Multi-Layer Perceptron (MLP).
  2. Deep Learning: Custom neural networks were built using TensorFlow/Keras, featuring 5 to 6 hidden layers with ReLU activations. Due to the extreme class imbalance (94% abnormal tissue), the authors utilized specific weight initializations based on logarithmic proportions to ensure the models could effectively learn from the minority "normal" class.

Key Results and Discussion

The experimental results revealed a significant performance gap between the two approaches, especially as the classification task became more nuanced.

Normal vs. Abnormal Tissue

Both methods performed well in distinguishing healthy brain tissue from tumorous tissue. Random Forest achieved a raw accuracy of 93.3%. However, the Deep Learning model provided superior Specificity (99.6%), which is arguably the more critical metric in surgery, as it minimizes the risk of false positives that could lead to the accidental removal of healthy brain tissue.

Low-Grade vs. High-Grade Malignancy

The most striking finding occurred in the grading task. Classical machine learning methods struggled to capture the subtle, non-linear metabolic differences between WHO Grade I-II (low-grade) and Grade III-IV (high-grade) tumors, with accuracies plateauing between 61% and 68.6%. In contrast, the Deep Learning models achieved 87.1% accuracy (84.7% balanced accuracy). This suggests that deep neural networks are uniquely capable of extracting complex features from the high-dimensional lipid signatures present in DESI-MSI data.

Conclusion

This research demonstrates that deep learning can bridge the gap between complex chemical data and real-time clinical decision-making. While challenges remain—such as addressing the scarcity of "normal" tissue samples and expanding the models to identify specific histological subtypes—the high accuracy and fast inference times of deep learning models suggest they could eventually provide neurosurgeons with high-precision, real-time guidance during tumor resection. Future work will focus on refining signal preprocessing and extending the classification to a wider variety of pediatric brain tumor types.

Summary updated

Abstract

Pediatric brain tumors are the most common cause of death among all childhood cancers and surgical resection usually is the first step in disease management. During surgery, it is important to perform safe gross resection of tumors, retaining as much brain tissue as possible. Therefore, appropriate resection margin delineation is extremely relevant. Currently available methods for tissue analysis have limited precision, are time-consuming, and often require multiple invasive procedures. Our main goal is to test whether machine learning techniques are capable of classifying the pediatric brain tissue chemical profile generated by DESI-MSI, which is mainly lipidic, into normal or abnormal tissue and into low- and high-grade malignancy subareas within each sample. Our experiments show that deep learning methods outperform classical machine learning methods in the task of classifying brain tissue from DESI-MSI mass spectra, both in normal versus abnormal tissue, and, for malignant tissues, in low-grade versus high-grade malignancy. Our conclusion are based on the analysis of 34,870 annotated spectra, obtained from the neoplastic and non-neoplastic microanatomical stratification of individual samples from 116 pediatric patients who underwent brain tumor surgical resection at the Boldrini Children’s Center between 2000 and 2020. Support Vector Machines, Random, Forests, and Least Absolute Shrinkage and Selection Operator (LASSO) were among the classical machine learning techniques evaluated. • It is feasible to train a learning model to help classify brain tissue. • For this task, deep learning methods perform better than classical ones. • Distinction between normal and abnormal tissue is easier to achieve than distinction between low- and high-grade tumor.
BibTeX
@inproceedings{bezerra2024deep,
  author = {Thais Maria Santos Bezerra and Matheus Silva de Deus and Felipe Cavalaro and Denise Ribeiro and Ana Luíza Seidinger and Izilda Aparecida Cardinalli and Andréia M. Porcari and Luciano de Souza Queiróz and Hélio Pedrini and João Meidânis},
  title = {{Deep learning outperforms classical machine learning methods in pediatric brain tumor classification through mass spectra}},
  booktitle = {Intelligence-Based},
  year = {2024},
  doi = {10.1016/j.ibmed.2024.100178},
}
Figure