2026 · Artificial Intelligence class
Leukemia classification neural network
The goal of this project was to make a CNN to classify blood samples into 4 categories of acute lymphoblastic leukemia. The dataset contains 4000 images, 1000 for each category. As a experiment I trained 2 models, a big and small models, to compare how accurate both models can be.
Built with
- TenserFlow
- Python
- Vue
- Colab
Working demo
Loading Model
Downloading weights...
1. Input Image
Drag photo here
or click to browse
2. Classify
3. Prediction
Result
...
Making the model
TensorFlow in Python was used to train the models and TensorFlow.js to run the small model in the browser. The starting layers pre-trained optimized for image classification EfficientNetB0 and B7. They were fine tuned during the training proces.
Before even starting to train the model you need to:
- Make a list of all training images and there labels.
- Make a dataset for training, validation and testing.
- Perform transformations on the training and validation datasets in order to get more accurate training
Classification report small model
| Precision | Recall | F1-Score | Support | |
|---|---|---|---|---|
| Benign | 1.00 | 1.00 | 1.00 | 36 |
| Early | 1.00 | 1.00 | 1.00 | 90 |
| Pre | 0.52 | 1.00 | 0.69 | 68 |
| Pro | 1.00 | 0.07 | 0.14 | 67 |
| Accuracy | 0.76 | 261 | ||
| Macro Avg | 0.88 | 0.77 | 0.71 | 261 |
| Weighted Avg | 0.88 | 0.76 | 0.70 | 261 |
model_small = tf.keras.applications.EfficientNetB0(
include_top=False,
weights="imagenet",
input_shape=img_shape
)
model_small = Sequential([
base_model,
GlobalAveragePooling2D(),
Dense(128, activation='relu'),
Dropout(0.2),
Dense(class_count, activation='softmax')
])
model_small.compile(
optimizer=Adam(learning_rate=0.001),
loss='categorical_crossentropy',
metrics=['accuracy']
)Posable improvements
Besides EfficientNetB0 and B7 TenserFlow offers many other pre-trained modes to use. This application can posable replaced by one of the following models:
Jovica.me
Contact me:
hi@jovica.me