Primary supervisor

Jesmin Nahar

Brain cancer—particularly High-Grade Gliomas (HGG) like Glioblastoma—remains one of the most lethal central nervous system malignancies, primarily due to non-specific early symptoms, rapid infiltrative growth, and late-stage clinical detection. Standard diagnostic pathways rely heavily on invasive biopsies and complex neuroimaging (e.g., MRI), which can delay intervention and lead to poor overall survival outcomes.

To overcome these critical barriers, recent research leverages Machine Learning (ML) paired with Explainable AI (XAI) to construct non-invasive, highly accurate risk-stratification and predictive frameworks for early detection and survival modeling.

1. Research Objectives & Clinical Rationale

  • The Challenge: Early-stage brain tumors frequently present with non-specific neurological symptoms (headaches, mild cognitive changes), causing delays in specialized neuro-oncological referral.

  • The Solution: Develop a non-invasive tool integrating liquid biopsy biomarkers (circulating tumor DNA, microRNAs), routine serum/inflammatory panels, and radiomic features extracted from standard MRI scans.

  • Goal: Enable early risk triage, accelerating diagnosis and surgical planning to maximize complete surgical resection and post-operative therapeutic response.

2. Dataset Preprocessing & Modeling Framework

  • Multi-Modal Data Integration: Predictive models combine structural radiomics (tumor volume, intensity, edema margins) with clinical variables, molecular markers (IDH mutation status, MGMT methylation), and peripheral blood indicators.

  • Data Preprocessing: Class imbalances between high-grade, low-grade, and non-malignant lesions are addressed using SMOTE, while missing dataset values are resolved via K-Nearest Neighbors imputation.

  • Algorithms & Performance: Ensemble machine learning models—specifically XGBoost, LightGBM, and Random Forest—demonstrate high diagnostic discrimination, consistently achieving predictive accuracies between 88% and 94% (AUC > 0.92).

3. Explainable AI (XAI) for Clinical Interpretability

To bridge the gap between AI models and neuro-oncologist trust, Explainable AI frameworks are integrated:

  • SHAP (Shapley Additive exPlanations): Ranks primary feature drivers across patient cohorts, identifying tumor volume, radiomic texture features, IDH status, and patient age as top predictive indicators.

  • LIME (Local Interpretable Model-agnostic Explanations): Delivers individual case explanations, providing clinicians with explicit breakdowns of why a specific patient was stratified into a high-risk or low-risk category.

4. Key Findings & Clinical Significance

  • Accessible Diagnostic Triage: Combining non-invasive radiomics with routine blood parameters offers a scalable screening tool to prioritize urgent MRI neuroimaging in primary care settings.

  • Advanced Intervention Window: Early risk stratification accelerates surgical and adjuvant therapy timelines, directly contributing to extended progression-free and overall survival rates.

Aim/outline

To develop an interpretable, non-invasive machine learning framework for early brain cancer detection and risk stratification using multi-modal clinical, radiomic, and molecular data.

Outline

  1. Introduction: Clinical challenges, late-stage diagnosis, and the need for AI-driven early detection.

  2. Data Preprocessing: Multi-modal data integration, missing value imputation, and class imbalance handling.

  3. Model Development: Feature selection and training ensemble models (XGBoost, LightGBM, Random Forest).

  4. Explainable AI (XAI): Applying SHAP and LIME for model transparency and clinical trust.

  5. Evaluation & Impact: Validation via AUC-ROC/sensitivity and clinical implications for improved patient survival.

Required knowledge

For IT students conducting research on brain cancer prediction, required knowledge spans data preprocessing, predictive modeling, explainable AI, and clinical evaluation: students must understand how to integrate multi-modal data (combining MRI radiomic features with clinical and molecular markers), handle class imbalances and missing values, implement ensemble models like XGBoost, LightGBM, and Random Forest, apply Explainable AI tools like SHAP and LIME to ensure model transparency, and evaluate performance using clinical metrics such as AUC-ROC, precision-recall, and sensitivity rather than standard accuracy alone.