CNN-TumorNet: leveraging explainability in deep learning for precise brain tumor diagnosis on MRI images

Frontiers in Oncology, Год журнала: 2025, Номер 15

Опубликована: Март 26, 2025

Introduction The early identification of brain tumors is essential for optimal treatment and patient prognosis. Advancements in MRI technology have markedly enhanced tumor detection yet necessitate accurate classification appropriate therapeutic approaches. This underscores the necessity sophisticated diagnostic instruments that are precise comprehensible to healthcare practitioners. Methods Our research presents CNN-TumorNet, a convolutional neural network categorizing images into non-tumor categories. Although deep learning models exhibit great accuracy, their complexity frequently restricts clinical application due inadequate interpretability. To address this, we employed LIME technique, augmenting model transparency offering explicit insights its decision-making process. Results CNN-TumorNet attained 99% accuracy rate differentiating from scans, underscoring reliability efficacy as instrument. Incorporating guarantees model’s judgments comprehensible, enhancing adoption. Discussion Despite overarching challenge interpretability persists. These may function ”black boxes,” complicating doctors’ ability trust accept them without comprehending rationale. By integrating LIME, achieves elevated alongside transparency, facilitating environments improving care neuro-oncology.

Язык: Английский

A Combined Approach Using T2*-Weighted Dynamic Susceptibility Contrast MRI Perfusion Parameters and Radiomics to Differentiate Between Radionecrosis and Glioma Progression: A Proof-of-Concept Study

Life, Год журнала: 2025, Номер 15(4), С. 606 - 606

Опубликована: Апрель 5, 2025

Differentiating tumor progression from radionecrosis in patients with treated brain glioma represents a significant clinical challenge due to overlapping imaging features. This study aimed develop and evaluate machine learning model that integrates radiomics features T2*-weighted Dynamic Susceptibility Contrast MRI perfusion (DSC MRI) parameters improve diagnostic accuracy distinguishing these entities. A retrospective cohort of 46 (25 confirmed radionecrosis, 21 progression) was analyzed. From lesion segmentation on DSC MRI, 851 were extracted using PyRadiomics, alongside seven (e.g., relative cerebral blood volume, time peak) obtained time–intensity curves (TICs). These combined into single dataset 14 classification algorithms evaluated GroupKFold cross-validation (k = 4). The top-performing selected based predictive area under the curve (AUC) yield. Logistic Regression classifier achieved highest performance, an AUC 0.88, followed by multilayer perceptron AdaBoost values 0.85 0.79, respectively. precision 72%, 74%, 78% for three models, respectively, while 63%, 70%, 71%. Key variables included like wavelet-HHH_firstorder_Mean mean normalized TIC values. Our approach integrating shows strong potential progression. However, further validation larger cohorts is essential confirm generalizability this approach.

Язык: Английский

Explainable deep stacking ensemble model for accurate and transparent brain tumor diagnosis

Computers in Biology and Medicine, Год журнала: 2025, Номер 191, С. 110166 - 110166

Опубликована: Апрель 17, 2025

Early detection of brain tumors in MRI images is vital for improving treatment results. However, deep learning models face challenges like limited dataset diversity, class imbalance, and insufficient interpretability. Most studies rely on small, single-source datasets do not combine different feature extraction techniques better classification. To address these challenges, we propose a robust explainable stacking ensemble model multiclass tumor that combines EfficientNetB0, MobileNetV2, GoogleNet, Multi-level CapsuleNet, using CatBoost as the meta-learner improved aggregation classification accuracy. This approach captures complex characteristics while enhancing robustness The proposed integrates CapsuleNet within framework, utilizing to improve We created two large by merging data from four sources: BraTS, Msoud, Br35H, SARTAJ. tackle applied Borderline-SMOTE augmentation. also utilized methods, along with PCA Gray Wolf Optimization (GWO). Our was validated through confidence interval analysis statistical tests, demonstrating superior performance. Error revealed misclassification trends, assessed computational efficiency regarding inference speed resource usage. achieved 97.81% F1 score 98.75% PR AUC M1, 98.32% 99.34% M2. Moreover, consistently surpassed state-of-the-art CNNs, Vision Transformers, other methods classifying across individual datasets. Finally, developed web-based diagnostic tool enables clinicians interact visualize decision-critical regions scans Explainable Artificial Intelligence (XAI). study connects high-performing AI real clinical applications, providing reliable, scalable, efficient solution

Язык: Английский