Scholarlink

Mental illnesses are a worldwide health concern impacting people from various cultures. For treatment and support to work, it is important to find the problem early and get the right diagnosis. This work focuses on diagnosing mental problems using DL algorithms applied to MRI scans. The dataset, collected via Kaggle, has 12,010 images divided into four mental illness classifications. The preprocessing pipeline comprises resizing, rescaling, denoising, histogram equalization, contrast improvement, and data augmentation. Several state-of-the-art Convolutional Neural Network (CNN) designs are investigated, including VGG16, VGG19, ResNet101, DenseNet201, NasNet, ShuffleNet, and EfficientNetB0. Hybrid models, such as VGG16-ResNet101 and VGG16-DenseNet201, are proposed to utilize the strengths of these architectures. The greatest results are reached with the VGG16-ResNet101 hybrid, attaining a classification accuracy of 99.21%.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning for Mental Health: Automated Detection of Mental Disorders from MRI Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Manoara Begum,
  • Nasim Haider,
  • Bijoy Dey,
  • Rayhanuzzaman,
  • Tanjim Mahmud,
  • Saparbayev Madraxim,
  • Rakhimjon Rajapboyevich Rakhimov,
  • Abubokor Hanip,
  • Mohammad Shahadat Hossain

摘要

Mental illnesses are a worldwide health concern impacting people from various cultures. For treatment and support to work, it is important to find the problem early and get the right diagnosis. This work focuses on diagnosing mental problems using DL algorithms applied to MRI scans. The dataset, collected via Kaggle, has 12,010 images divided into four mental illness classifications. The preprocessing pipeline comprises resizing, rescaling, denoising, histogram equalization, contrast improvement, and data augmentation. Several state-of-the-art Convolutional Neural Network (CNN) designs are investigated, including VGG16, VGG19, ResNet101, DenseNet201, NasNet, ShuffleNet, and EfficientNetB0. Hybrid models, such as VGG16-ResNet101 and VGG16-DenseNet201, are proposed to utilize the strengths of these architectures. The greatest results are reached with the VGG16-ResNet101 hybrid, attaining a classification accuracy of 99.21%.