Brain tumor diagnosis in MRI scans images using Residual/Shuffle Network optimized by augmented Falcon Finch optimization

Xiaohang Guo, Tianyi Liu, Qinglong ChiScientific 2024 · DOI

Abstract

Brain tumor diagnosis is an important task in prognosing and treatment planning of the patients with brain cancer. In the meantime, using Magnetic Resonance Imaging (MRI) as a commonly used non-invasive imaging technique provides experts a helpful view for detecting brain tumors. While deep learning methods have shown significant success in analyzing medical images, they often require careful design of architecture and tuning of hyperparameters to achieve optimal results. This study presents a new approach for diagnosing brain tumors in MRI scans using deep learning, focusing on Residual/Shuffle Networks. To enhance the proposed network for brain tumor classification, a modified metaheuristic algorithm named Augmented Falcon Finch Optimization (AFFO) is introduced. AFFO utilizes bio-inspired principles to effectively search for the best hyperparameter configurations, thereby enhancing the reliability and accuracy of the deep learning model. The performance of the proposed method is evaluated on a standard brain tumor MRI dataset and compared with existing techniques to illustrate the effectiveness of combining Residual/Shuffle Networks with AFFO.

Introduction

While Deep Learning (DL) has revolutionized medical imaging, two major hurdles remain: architectural complexity and hyperparameter sensitivity. Designing a neural network that is deep enough to capture intricate tumor features but efficient enough for clinical use is difficult. Furthermore, "tuning" variables like learning rate, batch size, and dropout usually requires exhaustive manual experimentation or grid searches that are computationally expensive. In the context of brain tumors (Glioma, Meningioma, and Pituitary tumors), misdiagnosis can lead to catastrophic delays in treatment. This research aims to bridge the gap between high-performance AI and practical clinical utility by creating a self-optimizing system that removes the guesswork from model training.

Methodology: Residual/Shuffle Networks and AFFO

The proposed system utilizes a two-pronged approach combining a novel architecture with a bio-inspired optimizer.

The Residual/Shuffle Network (RSN)

This architecture merges the strengths of two established DL concepts: Residual connections (from ResNet) to allow for deeper networks without the vanishing gradient problem, and Shuffle units (from ShuffleNet) to encourage efficient information flow across feature channels. This hybrid structure is designed to capture both local textures and global anatomical structures in MRI scans efficiently.

The hybrid Residual/Shuffle Network (RSN) architecture featuring channel splitting and shuffle units for efficient feature extraction.

Augmented Falcon Finch Optimization (AFFO)

AFFO is a metaheuristic algorithm inspired by the foraging and perching behaviors of finches and falcons. The authors augmented the base algorithm with Lévy Flight (a mathematical random walk to avoid local optima) and Self-adaptive Population Adjustment. AFFO was used to automatically find the best values for seven critical hyperparameters, including learning rate, number of residual blocks, and dropout rate.

Data Acquisition and Preprocessing

The study utilized a standard, publicly available Brain Tumor Dataset from Figshare, containing 3,064 T1-weighted contrast-enhanced MRI images from 233 patients. The dataset covers three tumor types: Glioma (1,426 images), Pituitary (930 images), and Meningioma (708 images).

A critical finding of the study was that preprocessing is non-negotiable for high accuracy. The authors implemented a rigorous pipeline involving skull stripping to remove non-brain tissue and data augmentation (rotations, flips, and cropping) to help the model generalize.

The skull-stripping preprocessing pipeline used to isolate brain tissue and improve model focus on tumor regions.

Key Results

The proposed method was compared against industry standards including ResNet, AlexNet, VGG-16, Inception V3, and U-Net. Key performance metrics include:

  • High Accuracy: The proposed model reached 98.8% validation accuracy and 97.5% test accuracy, significantly outperforming VGG-16 (93.94%).
  • Computational Efficiency: The model required only 1.01 × 10⁹ FLOPs per iteration, making it significantly lighter and faster than U-Net (3.12 × 10⁹) or Inception V3 (2.56 × 10⁹).
  • Rapid Convergence: The AFFO-optimized model reached 95% accuracy within just 200 iterations, whereas competing models often required 500 to 1,000 iterations.

Conclusion and Future Work

The combination of a hybrid RSN architecture and AFFO hyperparameter tuning provides a robust, efficient solution for brain tumor diagnosis. While the results are promising, future work will focus on integrating multi-modal data (T2, FLAIR) and improving performance on rarer, atypical tumor presentations. Additionally, reducing the GPU resource requirements remains a priority for deployment in resource-constrained clinical settings.

Summary updated

Abstract

Brain tumor diagnosis is an important task in prognosing and treatment planning of the patients with brain cancer. in the meantime, using the Magnetic Resonance Imaging (MRI) as a commonly used non-invasive imaging technique provide the experts a helpful view for detecting the brain tumors. While deep learning methods have shown significant success in analyzing medical images, they often require careful design of architecture and tuning of hyperparameters to achieve optimal results. This study presents a new approach for diagnosing brain tumors in MRI scans using deep learning, focusing on Residual/Shuffle Networks. The designed network structures offer efficient results when compared to traditional deep learning models. To enhance the proposed network for brain tumor classification, a modified metaheuristic algorithm named Augmented Falcon Finch Optimization (AFFO) is introduced. AFFO utilizes bio-inspired principles to effectively search for the best hyperparameter configurations, thereby enhancing the reliability and accuracy of the deep learning model. The performance of the proposed method is evaluated on a standard brain tumor MRI dataset and compared with existing techniques, including ResNet, AlexNet, VGG-16, Inception V3, and U-Net to illustrate the effectiveness of combining Residual/Shuffle Networks with AFFO for brain tumor diagnosis.
BibTeX
@inproceedings{guo2024brain,
  author = {Xiaohang Guo and Tianyi Liu and Qinglong Chi},
  title = {{Brain tumor diagnosis in MRI scans images using Residual/Shuffle Network optimized by augmented Falcon Finch optimization}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-77523-2},
}
Figure