Summary

I have built a multi-classifier on the HAM10000 dataset, which contains images of 7 types of skin cancer.

Due to the significant class imbalances, I applied data augmentation on the less represented classes of the dataset.

My model is based on DenseNet121, allowing trainablity on some of its layers. To tackle overfitting, I reconstructed the top by incorporating multiple dropout layers.

The model tested an accuracy of 86%

Detailed description

Below you can find a detailed decription of each step of the project, including motivations and code snippets.
To access full code, refer to the jupyter notebook in the repository.

Original data

The dataset is composed of roughly 10.000 images of skin lesions, with an highly unbalanced distribution between classes, as highlighted by the plot:

Data augmentation

Since classes are unbalanced, I decided to perform data augmentation on the less represented classes (with less than 1000 images) on the training set.

I have built an Augmenter object, which creates 3 modified images from each image in the selected classes.

The applied modifications are:

• Rotation --> randomly between -90 and 90 degrees
• Flip --> with a 50% probability
• Blur --> addition of random intensity of gaussian blur
• Contrast --> random contrast between 0.4 (darker) and 1.4 (lighter)

The augmentation process is completely reproducible thanks to seeds.

Data Preparation

To decrease memory consumption, dataset is pre-processed and loaded in batches of 32 images at each iteration.

Pre-processing includes:

• Resizing to shape 256,256,3 (RGB channels).
• Rescaling values between 0 and 1.
• Storing labels into one hot encoded vectors.

Dataset was split into 70% training, 20% validation and 10% test.

The model

As the dataset is relatively small, I decided to use the DenseNet121 pre-trained model to allow for a better performance. I allowed for retraining of the model layers (except feature extraction section, as they are used to detect low-level features in images). I have rebuilt the top of the model by introducing multiple dropout layers to address overfitting.

base_model = DenseNet121(weights='imagenet', include_top=False, input_shape=(256, 256, 3))
for layer in base_model.layers[:149]:
    layer.trainable = False
for layer in base_model.layers[149:]:
    layer.trainable = True

model = Sequential()
model.add(base_model)
model.add(Flatten())
model.add(BatchNormalization())
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.7))
model.add(BatchNormalization())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(BatchNormalization())
model.add(Dense(64, activation='relu'))
model.add(Dropout(0.3))
model.add(Dense(7, activation='softmax'))

Inspired by: https://bouzouitina-hamdi.medium.com/transfer-learning-with-keras-using-densenet121-fffc6bb0c233

Training

The model was trained using Adam as optimizer, with a learning rate of 0.001. To address overfitting, a learning scheduler with 0.7 as decay rate was employed.

Best model was saved by using early stopping whenever the validation accuracy was not increasing for 3 epochs.

As the model is still presenting signs of overfitting, the use of a wider search space for hyperparameter tuning is needed. As of now, due to the lack of computational resources, the model was tested only for (0.7,0.75,0.8,0.85,0.9) decay rates and (64,128) batch sizes.

Testing

Recorded performances on the test set:

• Test Accuracy: 85.99%
• Test Loss: 0.4413

Potential Improvements

• Enhanced data augmentation -> Employing a broader range of diverse and realistic augmentations could facilitate the model in generalizing more effectively and efficiently. To achieve this, one might consider incorporating a Generative Adversarial Network.
• Hyperparameter tuning -> Due to resource constraints, a comprehensive grid search for fine-tuning hyperparameters (dropout rate, learning rate, batch size, decay factor, number of frozen layers) was not feasible. Utilizing parallelization could offer a potential solution.
• Altered architecture -> Employing a model with reduced depth and complexity might potentially aid in mitigating overfitting. Further testing would be necessary to assess its effectiveness.