[Paper Review] Joint brain tumor segmentation from multi MR sequences through a deep convolutional neural network

Farzaneh Dehghani, Alireza Karimian+1 more|arXiv (Cornell University)|Mar 7, 2022

Brain Tumor Detection and Classification19 citations

TL;DR

This study proposes a deep convolutional neural network for joint brain tumor segmentation across multiple MR sequences (FLAIR, T1, T2, and T1 contrast-enhanced), leveraging multimodal input to improve accuracy. The model achieves state-of-the-art performance by fusing features from all sequences, with ablation studies identifying the optimal combination for highest segmentation accuracy.

ABSTRACT

Brain tumor segmentation is highly contributive in diagnosing and treatment planning. The manual brain tumor delineation is a time-consuming and tedious task and varies depending on the radiologists skill. Automated brain tumor segmentation is of high importance, and does not depend on either inter or intra-observation. The objective of this study is to automate the delineation of brain tumors from the FLAIR, T1 weighted, T2 weighted, and T1 weighted contrast-enhanced MR sequences through a deep learning approach, with a focus on determining which MR sequence alone or which combination thereof would lead to the highest accuracy therein.

Motivation & Objective

• To automate brain tumor segmentation in multi-contrast MRI to reduce reliance on time-consuming manual delineation.
• To evaluate the contribution of individual MR sequences (FLAIR, T1, T2, T1ce) to segmentation accuracy.
• To determine the optimal combination of MR sequences for maximizing segmentation performance.
• To develop a deep learning framework capable of joint segmentation across multiple MRI contrasts.
• To provide a robust, observer-independent solution for clinical treatment planning and diagnosis.

Proposed method

• A U-Net-based deep convolutional neural network is employed as the segmentation backbone.
• Multimodal MRI features from FLAIR, T1, T2, and T1 contrast-enhanced sequences are concatenated early in the network for joint feature learning.
• Skip connections are used to preserve spatial details and improve gradient flow during training.
• The network is trained end-to-end using a combination of binary cross-entropy and dice loss to optimize segmentation performance.
• Ablation studies are conducted to evaluate the impact of individual sequences and their combinations on segmentation accuracy.
• Data augmentation techniques are applied to improve generalization and reduce overfitting on limited medical imaging data.

Experimental results

Research questions

• RQ1 Which individual MR sequence contributes most significantly to accurate brain tumor segmentation?
• RQ2 What is the optimal combination of MR sequences for maximizing segmentation performance?
• RQ3 How does multimodal feature fusion improve segmentation accuracy compared to single-sequence approaches?
• RQ4 Does the proposed deep learning model outperform existing state-of-the-art methods on standard benchmark datasets?
• RQ5 How robust is the model to variations in tumor size, location, and intensity across different MRI contrasts?

Key findings

• The model achieves the highest Dice similarity coefficient (DSC) when all four MRI sequences (FLAIR, T1, T2, and T1ce) are used together.
• FLAIR and T1ce sequences contribute the most to tumor boundary detection, particularly for enhancing and necrotic regions.
• The inclusion of T2-weighted images improves detection of edema and non-enhancing tumor components.
• The ablation study confirms that using only a single sequence results in significantly lower DSC compared to multimodal fusion.
• The proposed method outperforms baseline models trained on single sequences, demonstrating the benefit of joint learning across contrasts.
• The model shows consistent performance across diverse tumor types and locations, indicating robustness to anatomical variability.