OkanNet: A Lightweight Deep Learning Architecture for Classification of Brain Tumor from MRI Images

This paper introduces OkanNet, a lightweight custom CNN architecture for brain tumor classification that, while achieving lower accuracy (88.10%) than the transfer learning-based ResNet-50 model (96.49%), offers a significantly faster training time (3.2x speedup), making it a practical alternative for resource-constrained mobile and embedded systems.

The Gist

Imagine you are a doctor trying to diagnose a patient's brain tumor. You have a stack of hundreds of MRI scans (like a thick photo album) and you need to find the tumor, identify what kind it is, and decide on a treatment. Doing this manually is exhausting, slow, and humans can make mistakes when they are tired.

This paper introduces a digital assistant (an AI) designed to help doctors with this job. The authors, Okan Uçar and Murat Kurt, built two different types of AI assistants and compared them to see which one is better.

Here is the breakdown of their "tug-of-war" between two different approaches:

The Two Contenders

1. The "Lightweight Sprinter": OkanNet

Think of OkanNet as a compact, fuel-efficient electric car.

• What it is: The authors built this AI from scratch, designing a custom "brain" (neural network) specifically for brain tumors.
• The Goal: To be fast, light, and cheap to run. It doesn't have a massive engine; it has just enough power to get the job done quickly.
• The Result: It finished the training (learning) in about 5 minutes. It was very accurate (88%), but not perfect.
• Best Use Case: Imagine you are in a remote village or using a portable MRI machine on a tablet. You don't have a supercomputer, but you need an answer right now. OkanNet is perfect for this because it runs fast on small devices.

2. The "Heavyweight Champion": ResNet-50

Think of ResNet-50 as a massive, high-performance Formula 1 race car.

• What it is: This isn't built from scratch. It's a famous, pre-trained AI that has already "read" millions of photos of cats, dogs, cars, and trees (a dataset called ImageNet). The researchers just taught it how to look at brains instead of cats.
• The Goal: To be the absolute most accurate diagnostician possible, regardless of how much time or power it takes.
• The Result: It took about 16 minutes to train (3 times longer than OkanNet), but it was incredibly accurate (96.5%). It made fewer mistakes than the lightweight car.
• Best Use Case: Imagine a big hospital with powerful servers. They have time to wait for the computer to think, and they need the highest possible certainty before performing surgery. ResNet-50 is the choice here.

The Big Showdown: Speed vs. Accuracy

The researchers put these two against each other using a dataset of over 7,000 MRI scans. Here is what they found:

• The Accuracy Race: The "Formula 1 car" (ResNet-50) won. It correctly identified the tumor type 96.5% of the time. The "electric car" (OkanNet) was good, but only got it right 88% of the time.
• The Speed Race: The "electric car" (OkanNet) won easily. It learned the task 3.2 times faster than the heavy model.
• The Mistake Analysis: Both models were great at spotting "No Tumor" (healthy brains). However, when the tumors looked very similar to each other (like Glioma vs. Meningioma), the heavy model was better at telling them apart. The lightweight model sometimes got confused between these two similar-looking enemies.

The "So What?" (Why does this matter?)

This paper isn't just about which AI is "better." It's about knowing which tool to use for the right job.

• If you need the absolute best diagnosis and have a powerful computer in a big hospital, use ResNet-50. It's the gold standard for precision.
• If you need a quick screening on a mobile device, a laptop, or in a place with limited electricity, use OkanNet. It's a "good enough" solution that is fast and doesn't require a supercomputer.

The Future

The authors are excited about the future. They want to:

1. Put their lightweight OkanNet onto actual mobile phones so doctors can scan patients anywhere.
2. Try to make OkanNet even smarter by teaching it more advanced physics concepts (like how light bounces off surfaces) to help it see the tumors more clearly.

In a nutshell: They built a fast, small AI and a slow, smart AI. The slow one is more accurate, but the fast one is ready to go anywhere. Both are valuable tools in the fight against brain tumors.

Technical Summary

1. Problem Statement

The diagnosis and treatment planning of brain tumors (Glioma, Meningioma, Pituitary, and healthy tissue) rely heavily on Magnetic Resonance Imaging (MRI). However, manual analysis by radiologists is:

• Time-consuming: Involves reviewing hundreds of slices per patient.
• Error-prone: Susceptible to human fatigue and oversight.
• Resource-intensive: Requires significant computational power for high-accuracy Deep Learning models, limiting their deployment on mobile or embedded medical devices.

The study aims to develop an automated Computer-Aided Diagnosis (CAD) system that balances high classification accuracy with computational efficiency, addressing the trade-off between model depth and inference/training speed.

2. Methodology

The authors conducted a comparative study using a dataset of 7,023 MRI images (T1, T2, and FLAIR sequences) across four classes: Glioma, Meningioma, Pituitary, and No Tumor. The data was preprocessed by resizing to 224×224 pixels, converting grayscale to 3-channel RGB, and applying data augmentation (rotation, flipping, translation).

Two distinct approaches were implemented and compared:

A. Custom Architecture: OkanNet

• Design Philosophy: A lightweight Convolutional Neural Network (CNN) designed from scratch to minimize parameters and hardware requirements.
• Structure:
  • Three consecutive Convolutional Blocks.
  • Convolution: 3×3 kernels with filter counts increasing from 16 to 32 to 64.
  • Activation: ReLU (Rectified Linear Unit) after each convolution.
  • Normalization & Pooling: Batch Normalization and 2×2 Max Pooling (stride 2) in each block.
  • Classification: Fully connected layers (128 neurons), ReLU, 50% Dropout (to prevent overfitting), and a final Softmax layer for 4-class output.
• Goal: Fast training and suitability for low-power devices.

B. Transfer Learning: ResNet-50

• Design Philosophy: Utilizing a pre-trained deep architecture to leverage features learned from the massive ImageNet dataset.
• Implementation:
  • Used the 50-layer ResNet-50 model.
  • Fine-tuning: The original 1000-class output layer was removed and replaced with a new fully connected layer for the 4 brain tumor classes.
  • Strategy: "Skip Connections" were retained to mitigate the vanishing gradient problem, allowing the network to learn residual functions.

Experimental Setup

• Hardware: NVIDIA GeForce RTX 2060 GPU.
• Software: MATLAB R2023b with Deep Learning Toolbox.
• Hyperparameters: 8 Epochs, Mini-batch size of 32, Learning rate of $10^{-4}$, SGD with Momentum (SGDM).

3. Key Contributions

1. OkanNet Architecture: The proposal of a novel, custom CNN specifically optimized for brain tumor classification that requires significantly fewer computational resources than standard deep networks.
2. Comprehensive Comparative Analysis: A rigorous side-by-side evaluation of a custom lightweight model against a state-of-the-art Transfer Learning model (ResNet-50) on the same dataset.
3. Trade-off Quantification: Explicit demonstration of the relationship between model depth, accuracy, and training time, providing guidelines for selecting models based on deployment constraints (e.g., server vs. mobile).
4. Open Source Implementation: Provision of complete MATLAB source code for data preprocessing, model training, and evaluation metrics.

4. Experimental Results

The models were evaluated using Accuracy, Precision, Recall, F1-Score, and Training Time.

Metric	OkanNet (Custom)	ResNet-50 (Transfer Learning)
Accuracy	88.10%	96.49%
Precision	0.877	0.963
Recall	0.872	0.962
F1-Score	0.875	0.962
Training Time	311 seconds (~5 min)	1000 seconds (~16 min)

Key Findings:

• Performance: ResNet-50 achieved superior accuracy (96.49%), meeting high clinical reliability standards. It converged rapidly due to pre-trained weights.
• Efficiency: OkanNet was 3.2 times faster in training time (311s vs. 1000s) while maintaining a competitive accuracy of 88.10%.
• Error Analysis: Both models excelled at distinguishing "No Tumor" (Healthy) cases (>98% success). The primary source of error for both was distinguishing between Glioma and Meningioma, which share similar radiological tissue characteristics. ResNet-50 showed higher sensitivity in these difficult cases due to its deeper layers.

5. Significance and Conclusion

The study highlights a critical decision matrix for medical AI deployment:

• High-Accuracy Scenario: For server-based clinical decision support where precision is paramount and computational cost is secondary, ResNet-50 is the preferred choice.
• Edge/Mobile Scenario: For portable MRI units, mobile diagnostics, or environments with limited hardware resources, OkanNet offers a viable, efficient alternative. It provides acceptable accuracy with significantly reduced latency and resource consumption.

Future Work:
The authors propose integrating modern hybrid architectures (e.g., EfficientNet), using Generative Adversarial Networks (GANs) to address class imbalance, and implementing the OkanNet architecture on mobile platforms for real-time inference. They also suggest exploring advanced preprocessing techniques involving Bidirectional Reflectance Distribution Function (BRDF) models to further enhance feature extraction.