Research Portfolio · Deep Learning & Medical Imaging

Ali Naderi

Master of Researcher in Mechatronics Engineering at Urmia University of Technology — specializing in deep learning for clinical decision support and brain tumor diagnosis.

Brain Tumor Classification Convolutional Neural Networks Transfer Learning Channel Attention Medical Imaging EfficientNet
View Publication See Results
98.16%
Accuracy · 4-Class
99.4%
Accuracy · Binary
2025
Wiley · Complexity
AN
Ali Naderi
Mechatronics Engineering Researcher
🏛Urmia University of Technology, Urmia, Iran
🔬Department of Mechatronics Engineering
📅Published 2025 · DOI: 10.1155/cplx/1644859
Python TensorFlow Keras CNN MRI Analysis CAD Systems

Ali Naderi is a researcher at the Department of Mechatronics Engineering, Urmia University of Technology. His work focuses on the intersection of deep learning, computer vision, and clinical decision-making, with a particular emphasis on automated diagnosis of brain tumors from MRI data.

His research addresses a critical challenge in neuro-oncology: manual classification of brain tumors is time-consuming and error-prone. By developing computer-aided diagnosis (CAD) systems powered by state-of-the-art deep learning architectures, his work aims to give clinicians a reliable second opinion, reduce diagnostic errors, and improve patient outcomes.

His most recent publication introduces a novel architecture combining EfficientNetB7 with a Channel Attention Module, achieving top-tier performance across multiple benchmark datasets and surpassing all prior state-of-the-art methods on the same datasets.

Featured Publication
Journal Article · 2025
Wiley · Complexity · Volume 2025 · Article ID 1644859
Convolutional Neural Network and Channel Attention Mechanism for Multiclass Brain Tumor Classification
Ali Naderi, Akbar Asgharzadeh-Bonab, Farid Ahmadi, Hashem Kalbkhani
📅 Received: 30 Aug 2024 · Accepted: 7 Mar 2025
🔓 Open Access · Creative Commons Attribution
🔗 doi.org/10.1155/cplx/1644859
This paper introduces a novel deep learning model for multiclass classification of brain tumors using MRI, combining a fine-tuned EfficientNetB7 CNN with a Channel Attention Module and an optimized fully connected classifier. Transfer learning is applied by freezing the initial four blocks of EfficientNetB7 while retraining the subsequent three blocks for domain-specific feature extraction. The attention module refines extracted feature maps by emphasizing tumor-relevant channels. Experimental results confirm superior performance over recent approaches on Brats-4C, Brats-2C-large, and Brats-2C-small datasets, achieving 98.16%, 99.4%, and 99.2% accuracy respectively — establishing a new state of the art on all three benchmarks.
Proposed Architecture
Three-Stage Deep Learning Pipeline

The model processes brain MR images through a sequential pipeline of feature extraction, attention-based refinement, and grid-search optimized classification.

Input
Brain MRI
224×224
→
Augmentation
Data Aug.
Noise · Flip · Rotate
→
Feature Extraction
EfficientNetB7
Transfer Learning
→
Feature Refining
Channel Attn.
CAM Module
→
Classification
FC Classifier
Grid Search
→
Output
4 Classes
Glioma · Mening. · Pit. · None
🧠
Component 01
EfficientNetB7 Feature Extractor
A fine-tuned EfficientNetB7 pretrained on ImageNet. The first four MBConv blocks are frozen to preserve general features, while blocks 5–7 are retrained to capture tumor-specific patterns from MR images using skip connections and global average pooling.
7 Blocks 438 Layers 66.7M Params MBConv
👁
Component 02
Channel Attention Module
A CAM layer placed after EfficientNetB7 that infers a 1D attention map across channels. It aggregates spatial context via average- and max-pooling, passes both through a shared MLP, and applies sigmoid gating to amplify clinically relevant features.
Shared MLP AvgPool + MaxPool Sigmoid CBAM-style
⚙️
Component 03
Optimized FC Classifier
A fully connected classifier with batch normalization, a 256-neuron ReLU dense layer (selected via grid search), 45% dropout for regularization, and a softmax output layer. Trained with SGDM optimizer and cross-entropy loss with L1/L2 regularization.
256 Neurons Dropout 0.45 BatchNorm Grid Search
Experimental Results
State-of-the-Art Performance

Evaluated on three BraTS benchmark datasets using 5-fold stratified cross-validation. The proposed method surpasses all prior methods tested on the same datasets.

98.16%
Brats-4C
4-Class Accuracy
99.4%
Brats-2C-large
Binary Accuracy
99.2%
Brats-2C-small
Binary Accuracy
99.8%
Brats-4C
F1-Score
Comparison with Prior Methods — Brats-4C Dataset
Method Architecture Accuracy
Kang et al. (2021)DenseNet-169 + ShuffleNet + MnasNet
91.58%
Irmak (2021)Designed CNN
92.66%
Demir & Akbulut (2022)R-CNN + SVM
96.60%
Shahin et al. (2023)MPCANet (PCANet+CNN)
94.02%
Proposed MethodEfficientNetB7 + CAM + FC (Ours)
98.16%
Get in Touch
Collaboration & Contact

Open to research collaborations, clinical partnerships, and academic discussions in the fields of medical imaging and deep learning.

📧
Author
Alinaderi119@gmail.com
Ali Naderi (Corresponding)
🏛
Institution
Urmia University of Technology
Urmia, Iran
📄
Full Paper
doi.org/10.1155/cplx/1644859
Wiley · Complexity 2025