[Paper Review] Explainable Cost-Sensitive Deep Neural Networks for Brain Tumor Detection from Brain MRI Images considering Data Imbalance

Md. Tanvir Rouf Shawon, G. M. Shahariar+2 more|arXiv (Cornell University)|Aug 1, 2023

Brain Tumor Detection and Classification10 citations

TL;DR

The paper proposes a pipeline using CNNs and pre-trained models (ResNet50, InceptionV3, EfficientNetB0, NASNetMobile) with cost-sensitive learning to detect brain tumors in MRI images, plus multiple explainability methods (Grad-CAM variants, LIME). It reports state-of-the-art accuracy on a balanced dataset and improved performance on imbalanced data with cost-sensitive variants.

ABSTRACT

This paper presents a research study on the use of Convolutional Neural Network (CNN), ResNet50, InceptionV3, EfficientNetB0 and NASNetMobile models to efficiently detect brain tumors in order to reduce the time required for manual review of the report and create an automated system for classifying brain tumors. An automated pipeline is proposed, which encompasses five models: CNN, ResNet50, InceptionV3, EfficientNetB0 and NASNetMobile. The performance of the proposed architecture is evaluated on a balanced dataset and found to yield an accuracy of 99.33% for fine-tuned InceptionV3 model. Furthermore, Explainable AI approaches are incorporated to visualize the model's latent behavior in order to understand its black box behavior. To further optimize the training process, a cost-sensitive neural network approach has been proposed in order to work with imbalanced datasets which has achieved almost 4% more accuracy than the conventional models used in our experiments. The cost-sensitive InceptionV3 (CS-InceptionV3) and CNN (CS-CNN) show a promising accuracy of 92.31% and a recall value of 1.00 respectively on an imbalanced dataset. The proposed models have shown great potential in improving tumor detection accuracy and must be further developed for application in practical solutions. We have provided the datasets and made our implementations publicly available at - https://github.com/shahariar-shibli/Explainable-Cost-Sensitive-Deep-Neural-Networks-for-Brain-Tumor-Detection-from-Brain-MRI-Images

Motivation & Objective

• Motivate automated brain tumor detection to reduce manual review time in MRI reports.
• Evaluate CNN and pre-trained CNN architectures for brain tumor classification on balanced and imbalanced datasets.
• Incorporate cost-sensitive learning to handle data imbalance and compare with standard models.
• Apply multiple explainable AI methods to visualize and interpret model decisions.

Proposed method

• Evaluate CNN, ResNet50, InceptionV3, EfficientNetB0, and NASNetMobile for brain tumor detection on brain MRI datasets.
• Apply cost-sensitive learning to address data imbalance and compare with standard training.
• Fine-tune/pre-train models and employ a 224x224 input size with 80/10/10 train/val/test split.
• Use Grad-CAM, Grad-CAM++, Saliency Map, SmoothGrad, Score-CAM, Faster Score-CAM for gradient-based explanations and LIME for model-agnostic explanations.
• Assess performance with accuracy, precision, recall, F1-score, and specificity.
• Provide data preprocessing and model architecture details including dropout and learning rates as part of experimental settings.

Experimental results

Research questions

• RQ1 How do standard CNNs and pre-trained CNNs perform on balanced brain MRI datasets for tumor detection?
• RQ2 Can cost-sensitive learning improve performance on imbalanced brain tumor datasets compared to standard training?
• RQ3 How effective are explainable AI techniques (Grad-CAM, LIME, etc.) at interpreting brain tumor predictions from these models?
• RQ4 What is the impact on accuracy/recall when applying cost-sensitive approaches to CNNs like InceptionV3 and CNN?

Key findings

• InceptionV3 achieves 99.33% accuracy on the balanced dataset after fine-tuning.
• Cost-sensitive InceptionV3 (CS-InceptionV3) achieves 92.31% accuracy and CS-CNN achieves 92.31% accuracy with a recall of 1.00 on the imbalanced dataset.
• Cost-sensitive approaches provide about 4% higher accuracy on imbalanced data compared to conventional models used in the study.
• Explainability techniques (Grad-CAM, Grad-CAM++, Saliency Map, SmoothGrad, Score-CAM, Faster Score-CAM) were applied to interpret CNNs and LIME to explain pre-trained models.