Optimized U-Net Architecture with Enhanced Chimp Optimization Algorithm for Brain Tumor Segmentation

Akash Verma

doi:10.1007/s12559-025-10530-1

Return

Optimized U-Net Architecture with Enhanced Chimp Optimization Algorithm for Brain Tumor Segmentation

2025-12-05

PRE

Akash Verma *

Arun Kumar Yadav

DOI：10.1007/s12559-025-10530-1

Original

Original request for help

Save

Abstract

En 中文

Early detection and precise diagnosis of brain tumors are pivotal for enhancing patient outcomes. Magnetic Resonance Imaging (MRI) is commonly used for brain tumor identification, but manual detection from vast MRI datasets relies on medical specialist’s time and expertise. Deep learning, particularly convolutional neural networks, has excelled in medical image analysis. In the field of deep learning-driven medical image analysis, accurately identifying the tumorous region from healthy tissues is still an incredible open challenge. The commonly employed U-Net architecture has limitations in capturing a wide range of features because of its uniform encoder design and the stacking of convolutions. These limitations include excessive parameter count, high memory consumption, a deficiency in global context, incomplete segmentation, and imprecise identification of low-grade tumor boundaries. To tackle these challenges, we introduce an innovative architecture based on Nested U-Net, incorporating both up-sampling and down-sampling operations, as well as respath connection & multiple skip connections, to segment the brain tumors using 3-dimensional (3D) MRI images. Additionally, we optimize the hyperparameters using an Enhanced chimp optimization algorithm (ECHOA) method to maximize the Dice score achieved by our proposed model. Our proposed model demonstrates better Dice Scores across various tumor subregions (Whole tumor, Tumor Core, and Enhancing tumor) using Brain Tumor Segmentation (BraTS) 2017, 2018, 2019, 2020, and 2021 datasets. To improve the generosity of the model, the proposed model is cross validated using 5-fold, Leave-2-out and Leave-(n-1)-out approaches. The outcomes demonstrate that the proposed method surpasses existing state-of-the-art approaches across all the referenced datasets.

Keywords:

Deep learning

Medical image processing

Encoder-decoder

Brain tumor segmentation

Convolutional neural network