🚀 5 Mistakes I Made in My First CNN Project
machinelearning #deeplearning #cnn #ai
😅 I Thought My Model Was Working... Until I Looked Closer
When I built my first CNN model for brain tumor classification using MRI images, I initially felt pretty confident.
The code was running.
The model was training.
The accuracy looked reasonable.
And predictions were being generated.
The model was classifying images into:
- Glioma
- Meningioma
- Pituitary
- No Tumor
So I assumed everything was working.
Then I started looking beyond the overall accuracy.
That's when I realized:
A model can appear to work while still having important problems underneath.
Here are five mistakes that taught me more than simply getting the model to train.
❌ Mistake 1: Not Understanding Class Distribution
One of the first things I should have checked more carefully was the distribution of images across the four classes.
I should have asked:
- How many images are available for each class?
- Are the classes reasonably balanced?
- Are some classes significantly underrepresented?
- Could class distribution affect the metrics I'm seeing?
Instead, I focused on getting the CNN running.
What I learned
If one class has substantially more samples than another, overall accuracy may not tell the complete story.
A model can perform well overall while performing poorly on a less represented class.
Lesson
Understand the distribution of your data before interpreting your model's performance.
❌ Mistake 2: Increasing Model Complexity Without a Clear Reason
I initially had a simple assumption:
"More layers should give the model more ability to distinguish all four classes."
So I experimented with increasing the model's complexity.
The problem wasn't experimenting.
The problem was adding complexity without a clear hypothesis about what I was trying to improve.
What I observed
- Training performance improved
- Validation performance did not improve accordingly
- The gap between training and validation became more noticeable
That was a sign that the model could fit the training data better without necessarily generalizing better.
Lesson
A more complex model isn't automatically a better model.
Architecture changes should have a reason behind them.
❌ Mistake 3: Trusting Overall Accuracy Too Much
This was probably one of my biggest mistakes.
When I saw a decent overall accuracy, my first reaction was:
"The model is working well."
But I wasn't asking enough questions about how that accuracy was achieved.
I needed to look at things such as:
- Class-wise performance
- Confusion matrix
- Precision
- Recall
- F1-score
- Which classes were being confused
For example, the model showed confusion between some of the tumor categories.
That made me realize that:
Overall accuracy is only one view of model performance.
Lesson
For a multi-class classification problem, don't stop at:
"What percentage did the model get correct?"
Also ask:
"Which classes is the model getting wrong, and how often?"
❌ Mistake 4: Choosing Hyperparameters Without Understanding Them
Another mistake was treating hyperparameters almost like fixed values that could simply be copied from another implementation.
I experimented with things such as:
- Learning rate
- Batch size
- Number of epochs
- Model architecture
But initially, I didn't understand enough about why I was choosing those values.
What changed
I started treating each experiment as something I should be able to explain.
Instead of:
"This configuration worked for someone else, so I'll use it."
I started thinking:
"What am I changing, why am I changing it, and what effect should I expect?"
Lesson
Don't copy hyperparameters blindly. Understand what you're changing and evaluate the result.
There isn't one universally correct learning rate, batch size, or number of epochs for every CNN project.
❌ Mistake 5: Not Visualizing the MRI Data Early Enough
This sounds simple, but it was one of the most useful lessons.
I should have spent more time looking at the actual images before focusing on the model.
I wanted the CNN to discover the patterns.
But first, I needed to understand the data myself.
I started looking more closely at:
- Differences between classes
- Image quality
- Visual patterns
- Tumor locations
- Similarities between categories
🧠 What I Should Have Observed Earlier
Some categories can contain visually similar patterns.
That doesn't mean a human can simply look at an MRI and make the correct classification.
It means that the visual characteristics of the dataset matter when you're trying to understand where a model may struggle.
Lesson
Visualize your data before asking your model to learn from it.
🧠 What Changed After These Mistakes
The biggest change wasn't one particular model architecture.
It was my approach to experimentation.
I started doing more of this:
Understand → Experiment → Evaluate → Question → Improve
Instead of:
Build → Train → Check Accuracy → Done
I became more interested in questions like:
- Why did validation performance change?
- Which class is causing the problem?
- Is the model overfitting?
- Did the preprocessing actually help?
- Did the architectural change improve generalization?
- What does the confusion matrix tell me that accuracy doesn't?
🎯 My Biggest Takeaway
Multi-class classification isn't simply:
"Binary classification + more classes."
Each additional class introduces more opportunities for confusion and makes evaluation more important.
And perhaps the most useful lesson from this project was:
A model doesn't become reliable just because it trains successfully.
Training is only the beginning.
🔗 Part of My CNN Learning Series
This is the second post in my CNN learning series.
🧠 Part 1 — What My First CNN Project Taught Me
⚠️ Part 2 — 5 Mistakes I Made in My First CNN Project
You're reading this one.
⚙️ Part 3 — How I Built the CNN Model
Coming next.
🔍 Part 4 — Understanding Model Predictions with Grad-CAM
Coming later.
🔗 Project Reference
The project discussed in this article is part of my brain tumor detection and classification project.
The project includes:
- MRI image preprocessing
- CNN-based classification
- Binary and multi-class experiments
- Model evaluation
- Grad-CAM explainability
- Streamlit interface
Note: This project is an academic/learning project and is not intended to replace professional medical diagnosis.
👋 Let's Connect
I'm continuing to learn machine learning by building practical projects and documenting what I learn along the way.
If you're also working with CNNs, computer vision, or deep learning, I'd be interested to hear about your experience.
🌐 Portfolio
💻 GitHub
💼 LinkedIn
Thanks for Reading!
Top comments (0)