Logo image
Back
Shapley Additive exPlanations for Identification of Astrocytoma and Glioblastoma using Region-Based MRI Radiomics Feature Analysis
Conference proceeding

Shapley Additive exPlanations for Identification of Astrocytoma and Glioblastoma using Region-Based MRI Radiomics Feature Analysis

Md. Ashik Sarker Lifat, Md. Shahariar Sarkar, Kazi Al Ashfaq, Md. Mahedi Hassan, Rehana Karim and Rakibul Hasan
2025 2nd International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM), pp 1-4
27 Jun 2025
DOI: https://doi.org/10.1109/NCIM65934.2025.11160025

Abstract

3D MRI Accuracy Additives Brain Tumor Data models Feature extraction Radiomics SHAP Solid modeling Three-dimensional displays Training Brain Cancer Machine Learning Magnetic Resonance Imaging
Astrocytoma and Glioblastoma are two types of brain cancer also known as adult diffuse gliomas. These diseases are primarily characterized by histological criteria and molecular markers. Identifying these cancers by molecular diagnostics remains an invasive and time-consuming process. Previous studies have shown that radiomics features, especially from MR imaging have the potential to identify cancer-type characteristics accurately. However, tumor regional radiomics feature selection from MRI imaging for improved outcomes is still challenging. In this study, texture, shape, and 1st order features are extracted from the three parts of the tumor images. The filter-based feature selection pipeline is implemented to gain significant features from each region. Then, the selected features are combined for machine learning model development. The dataset is balanced with a split using a 5-fold cross-validation, and each fold of the training dataset SMOTE technique is used to balance the class. The data are normalized by the Robust scaler method and then trained using three machine learning models. The result shows that necrotic core (NC) and enhancing tumor (ET) regional features have more significance by mutual information scores and statistical tests. Random Forest (RF) classifier has performed better for both classes with an average accuracy of 88%, f1 score of 71%, and AUC of 83%. Finally, Shapley Additive Explanations (SHAP) are employed to interpret the input features that contribute to the outcomes of models. This result indicates that regional radiomics features with machine learning assist in the automated diagnosis of these diseases.

Metrics

1 Record Views

Details

Logo image