ith-project/skin_cancer_classification.py

# -*- coding: utf-8 -*-
"""Skin Cancer Classification.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1Nhk-lK1OCihOAxo36qt2e1_WgdLc-EiR

# Skin Cancer Classification
**Output**: Benign / Malignant

**Model**: DenseNet121 (transfer learning)

**Dataset**: Skin Cancer: HAM10000 dataset https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-ham10000/data

| Class | Name                                | Value     |
|-------|-------------------------------------|-----------|
| nv    | Melanocytic nevi (moles)            | Benign    |
| bcc   | Basal cell carcinoma                | Malignant |
| bkl   | Benign keratosis-like lesions       | Benign    |
| df    | Dermatofibroma                      | Benign    |
| vasc  | Vascular lesions                    | Benign    |
| mel   | Melanoma                            | Malignant |
| akiec | Actinic keratoses / Bowen's disease | Benign    |

## 1. Requirements and dataset download
"""

import sys
IN_COLAB = 'google.colab' in sys.modules

# if IN_COLAB:
#   !pip install pandas numpy matplotlib seaborn pillow scikit-learn tensorflow
#   !pip install --upgrade kagglehub[pandas-datasets,hf-datasets]

import kagglehub

# Download latest version
path = kagglehub.dataset_download("kmader/skin-cancer-mnist-ham10000")

print("Path to dataset files:", path)

# !cp -R /kaggle/input/skin-cancer-mnist-ham10000 /content/skin-cancer-mnist-ham10000

"""## 2. Imports and setup"""

import os
import glob
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

from PIL import Image

# Source - https://stackoverflow.com/a/53586419
# Posted by korakot, modified by community. See post 'Timeline' for change history
# Retrieved 2026-05-21, License - CC BY-SA 4.0

"""### Loading dataset

We start loading the CSV dataset into a dataframe:
"""

# Path to your dataset folder
dataset_path = path

# Metadata file
metadata_path = os.path.join(dataset_path, "HAM10000_metadata.csv")

# Load CSV
df = pd.read_csv(metadata_path)

# Show first rows
print(df.head())

"""Here we load all JPG images (path format) from the dataset to the dataframe:"""

# Collect all image paths
image_paths = glob.glob(os.path.join(dataset_path, "**", "*.jpg"), recursive=True)

# Create dictionary:
# key = image_id
# value = full image path
imageid_path_dict = {
    os.path.splitext(os.path.basename(x))[0]: x
    for x in image_paths
}

# Add image path column
df['path'] = df['image_id'].map(imageid_path_dict.get)

# Check
print(df[['image_id', 'path']].head())

"""## 3. Dataset analysis

### Missingness

The dataset does not have any missing cell:
"""

missing = df['path'].isnull().sum()

print(f"Missing images: {missing}")

"""### Class distribution

If we analyze the distribution of lesion type, we find that there is a clear unbalance. _Melanocytic nevi_ is the most abundant class.
"""

plt.figure(figsize=(10,5))

sns.countplot(data=df, x='dx', order=df['dx'].value_counts().index)

plt.title("Lesion type distribution")
plt.xlabel("Lesion type")
plt.ylabel("Count")
plt.show()

"""### Visualize samples

We visualize some random samples of the dataset, adding on the top the lesion type they belong:
"""

fig, axes = plt.subplots(2, 4, figsize=(10,5))

for i, ax in enumerate(axes.flat):
    sample = df.sample(1).iloc[0]

    img = Image.open(sample['path'])

    ax.imshow(img)
    ax.set_title(sample['dx'])
    ax.axis('off')

plt.tight_layout()
plt.show()

"""### Data distribution

It is interesting to study the distribution of other features in the dataset, in order to detect outliers.

We start by the age, plotting it in 20 bins. First we find that most abudnant age is arround 50 years old (from 40 to 58). One surprise is that we detect zero samples in two bins, which correspond to the ages from 55 to 60 years and 25 to 30 years old.
"""

plt.figure(figsize=(8,5))

sns.histplot(df['age'].dropna(), bins=20)

plt.title("Age Distribution")
plt.show()

"""However, sex distribution is quite balanced. Even so, there are more males than females."""

sns.countplot(data=df, x='sex')

plt.title("Sex Distribution")
plt.show()

"""Lesion localization is not equal probable in all parts of the body. Instead, back and lower extremity are the most common for skin lesions."""

plt.figure(figsize=(12,5))

sns.countplot(
    data=df,
    x='localization',
    order=df['localization'].value_counts().index
)

plt.xticks(rotation=45)
plt.title("Lesion Localization")
plt.show()

"""### Image sizes

The size of the images is 600 x 450 pixels.
"""

sizes = []

for path in df['path'].sample(100):
    img = Image.open(path)
    sizes.append(img.size)

print(pd.Series(sizes).value_counts())

"""## 2. Prepare dataset

In this section, we process the previous dataframe to make it valid for the model training.

### Encoding Binary Labels

Our goal is to classify between benign and malignant. Then, we should encode the lesion type to a new feature describing both states. This new feature is numerical: 0 means benign and 1, malignant.
"""

# Mapping from dx to benign/malignant
# 0 --> benign
# 1 --> malignant
benign_malignant_dict = {
    'nv': 0,
    'bcc': 1,
    'bkl': 0,
    'df': 0,
    'vasc': 0,
    'mel': 1,
    'akiec': 0
}

# Create new column
df['target'] = df['dx'].map(benign_malignant_dict)

# Preview
print(df[['dx', 'target']].head())

"""### Dataset split

We split the dataset into training (70%) and validation (15%) and test (15%) sets.
"""

from sklearn.model_selection import train_test_split

train_val_df, test_df = train_test_split(
    df,
    test_size=0.15,
    random_state=42,
    stratify=df['target']
)

train_df, val_df = train_test_split(
    train_val_df,
    test_size=0.1765,   # ≈ 15% of total dataset
    random_state=42,
    stratify=train_val_df['target']
)

print("Train size:", len(train_df))
print("Validation size:", len(val_df))
print("Test size:", len(test_df))

"""We convert to string format the encoded column for ImageDataGenerator."""

# Convert 'target' column to string type for ImageDataGenerator
train_df['target'] = train_df['target'].astype(str)
val_df['target'] = val_df['target'].astype(str)

print("Train target dtype after conversion:", train_df['target'].dtype)
print("Validation target dtype after conversion:", val_df['target'].dtype)

"""### Balancing classes by oversampling the minority class

We decide to handle class unbalance through three approaches:
* Minority class oversampling
* Class weight
* Data augmentation
"""

## Identify majority and minority classes in the training set
#class_counts = train_df['target'].value_counts()
#majority_class = class_counts.idxmax()
#minority_class = class_counts.idxmin()

## Get the DataFrames for majority and minority classes
#df_majority = train_df[train_df['target'] == majority_class]
#df_minority = train_df[train_df['target'] == minority_class]

## Oversample the minority class
#df_minority_oversampled = df_minority.sample(
#    class_counts[majority_class], replace=True, random_state=42
#)

## Combine majority class with oversampled minority class
#train_df_balanced = pd.concat([df_majority, df_minority_oversampled])

## Shuffle the balanced DataFrame
#train_df_balanced = train_df_balanced.sample(frac=1, random_state=42).reset_index(drop=True)

#print("Original train_df class distribution:")
#print(train_df['target'].value_counts())
#print("\nBalanced train_df_balanced class distribution:")
#print(train_df_balanced['target'].value_counts())

plt.figure(figsize=(6,3))

sns.countplot(data=train_df, x='target')

plt.title("Benign vs Malignant Distribution")
plt.xlabel("Lesion Type")
plt.ylabel("Count")
plt.show()

print(train_df['target'].value_counts())

#plt.figure(figsize=(6,3))

#sns.countplot(data=train_df_balanced, x='target')

#plt.title("Benign vs Malignant Distribution")
#plt.xlabel("Lesion Type")
#plt.ylabel("Count")
#plt.show()

#print(train_df_balanced['target'].value_counts())

"""### Class weight (class imbalance)"""

from sklearn.utils.class_weight import compute_class_weight

classes = np.unique(train_df['target'])

weights = compute_class_weight(
    class_weight='balanced',
    classes=classes,
    y=train_df['target']
)

class_weights = dict(enumerate(weights))

print(class_weights)

"""## 3. Build the model"""

import tensorflow as tf
from tensorflow.keras import layers, models
from tensorflow.keras.applications import DenseNet121
from tensorflow.keras.applications.densenet import preprocess_input
from tensorflow.keras.layers import Dense, GlobalAveragePooling2D
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam

gpus = tf.config.list_physical_devices('GPU')
print("GPUs:", gpus)

strategy = tf.distribute.MirroredStrategy()

data_augmentation = tf.keras.Sequential([
    tf.keras.layers.RandomFlip("horizontal"),
    tf.keras.layers.RandomRotation(0.1),
    tf.keras.layers.RandomZoom(0.1),
    tf.keras.layers.RandomContrast(0.1),
])

base_model = DenseNet121(
    weights='imagenet',
    include_top=False,
    input_shape=(224, 224, 3)
)

inputs = tf.keras.Input(shape=(224,224,3))

x = data_augmentation(inputs)

x = base_model(x)
x = GlobalAveragePooling2D()(x)
x = Dense(512, activation='relu')(x) # Added another Dense layer
x = Dense(256, activation='relu')(x) # Existing Dense layer
predictions = Dense(1, activation='sigmoid')(x) # Output layer for binary classification

model = Model(inputs=inputs, outputs=predictions)
model.compile(optimizer=Adam(learning_rate=0.0001), loss='binary_crossentropy', metrics=['accuracy'])

"""## 4. Data Generators

I will prepare data generators for training and validation. The training generator will include data augmentation and preprocessing, while the validation generator will only preprocess the images.
"""

from tensorflow.keras.preprocessing.image import ImageDataGenerator

# Image dimensions
IMG_WIDTH = 224
IMG_HEIGHT = 224

# Data generators
train_datagen = ImageDataGenerator(
    preprocessing_function=preprocess_input,
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
    fill_mode='nearest'
)

val_datagen = ImageDataGenerator(preprocessing_function=preprocess_input)

# Flow from dataframe
train_generator = train_datagen.flow_from_dataframe(
    dataframe=train_df,
    x_col='path',
    y_col='target',
    target_size=(IMG_WIDTH, IMG_HEIGHT),
    batch_size=32,
    class_mode='binary',
    seed=42
)

val_generator = val_datagen.flow_from_dataframe(
    dataframe=val_df,
    x_col='path',
    y_col='target',
    target_size=(IMG_WIDTH, IMG_HEIGHT),
    batch_size=32,
    class_mode='binary',
    seed=42
)

"""Now, the `train_generator` will use the `train_df_balanced` DataFrame, which has an equal number of samples for both classes. This will help the model learn more effectively from the minority class during training.

## 6. Train the Model
"""

from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint

model_checkpoint = ModelCheckpoint(
    'best_model.keras',
    monitor='val_accuracy',
    save_best_only=True,
    mode='max'
)

# Train the model
history = model.fit(
    train_generator,
    epochs=30, # You can adjust the number of epochs
    validation_data=val_generator,
    callbacks=[model_checkpoint],
    class_weight=class_weights # Use class weights to handle imbalance
)

"""## 7. Evaluation"""

IMG_WIDTH = 224
IMG_HEIGHT = 224

test_datagen = ImageDataGenerator(
    preprocessing_function=preprocess_input
)

test_generator = test_datagen.flow_from_dataframe(
    dataframe=test_df,
    x_col='path',
    y_col='target',
    target_size=(IMG_WIDTH, IMG_HEIGHT),
    batch_size=32,
    class_mode='binary',
    shuffle=False   # CRITICAL
)

"""### Load best model"""

from tensorflow.keras.models import load_model

best_model = load_model('best_model.keras')

"""### Evaluate on test set"""

results = best_model.evaluate(test_df)

for name, value in zip(best_model.metrics_names, results):
    print(f"{name}: {value:.4f}")

"""### Predictions and Classification Report"""

from sklearn.metrics import classification_report, confusion_matrix

test_generator.reset()
y_prob = best_model.predict(test_generator)
y_pred = (y_prob > 0.5).astype(int).ravel()

y_true = val_generator.classes

print("Classification Report:")
print(classification_report(y_true, y_pred))

"""### Confusion Matrix"""

from sklearn.metrics import confusion_matrix
import seaborn as sns
import matplotlib.pyplot as plt

cm = confusion_matrix(y_true, y_pred)

plt.figure(figsize=(6,5))

sns.heatmap(
    cm,
    annot=True,
    fmt='d',
    cmap='Blues',
    xticklabels=["Benign", "Malignant"],
    yticklabels=["Benign", "Malignant"]
)

plt.xlabel("Predicted")
plt.ylabel("True")
plt.title("Test Confusion Matrix")

plt.show()

"""### Classification report"""

from sklearn.metrics import classification_report

print(classification_report(
    y_true,
    y_pred,
    target_names=["Benign", "Malignant"]
))

"""### ROC-AUC"""

from sklearn.metrics import roc_auc_score

auc = roc_auc_score(y_true, y_prob)

print(f"ROC-AUC: {auc:.4f}")

"""### Recall"""

from sklearn.metrics import recall_score

sensitivity = recall_score(y_true, y_pred)

print(f"Sensitivity: {sensitivity:.4f}")

"""### Training History Plots"""

plt.figure(figsize=(12, 5))

# Plot training & validation accuracy values
plt.subplot(1, 2, 1)
plt.plot(history.history['accuracy'])
plt.plot(history.history['val_accuracy'])
plt.title('Model Accuracy')
plt.ylabel('Accuracy')
plt.xlabel('Epoch')
plt.legend(['Train', 'Validation'], loc='upper left')

# Plot training & validation loss values
plt.subplot(1, 2, 2)
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('Model Loss')
plt.ylabel('Loss')
plt.xlabel('Epoch')
plt.legend(['Train', 'Validation'], loc='upper left')

plt.tight_layout()
plt.show()