Hours:
8 hours (2 credits)

Room:

Aula Riunioni del Dipartimento di Ingegneria dell’Informazione, Via G. Caruso 16, Pisa - Ground Floor

To register to the course, click here

Short Abstract:

This 8-hour course will cover a comprehensive exploration of machine learning techniques applied to medical imaging, leading towards the early detection of melanoma. The course will cater to a diverse audience, from beginners to intermediate practitioners in machine learning, and will blend theoretical foundations with practical applications.

The course will begin with an overview of the iToBoS EU project, emphasizing the significant role of machine learning in enhancing the accuracy and efficiency of melanoma detection. This will set the stage for the technical content that follows, and highlighting the real-world impact of these technologies in healthcare.

Participants will then be introduced to the foundational concepts of machine learning, starting with linear regression, and their application in predictive analytics. The course will then transition into logistic regression, focusing on classification problems relevant to medical diagnostics.

As the course progresses, it will delve into neural networks, using logistic regression as a stepping stone to understanding neural networks. The progression from logistic regression to neural networks is a natural evolution from simpler linear models to more complex, non-linear models capable of handling a wider range of data and problem types in machine learning.

The course will culminate in an in-depth exploration of Convolutional Neural Networks (CNNs), vital for image recognition and analysis tasks in melanoma detection. Attendees will learn about the architecture of CNNs, their layers and filters, and the application of CNNs to dermoscopic images for melanoma detection.

Furthermore, the course will address the challenge of working with unbalanced datasets in medical imaging. It will introduce specialized evaluation metrics such as Precision, Recall, F1 Score, AUC-ROC, and the Precision-Recall Curve, which are essential for accurately assessing model performance in scenarios where traditional metrics may be misleading.

By the end of this course, participants will have gained a solid understanding of key machine learning concepts and their application in the medical field, particularly in melanoma detection.

Course Contents in brief:

Lecture 1: Learning Models (2 hours)

1. Overview of the iToBoS project
   • Early detection of melanoma
2. Supervised vs Unsupervised Learning
   • Understanding the structure of the data
   • Differentiating between regression and classification.
3. Predictive modeling: Linear Regression
   • Basic concept.
4. Modeling the hypothesis function
   • Formulating the hypothesis function.
   • Understanding linear regression parameters.
5. Cost Function and Gradient Descent
   • Mean Squared Error cost function
   • Implementing gradient descent.

Lecture 2: Logistic Regression (1 hour)

1. A linear classifier
   • Understanding logistic regression for classification.
   • Binary classification examples.
2. Hypothesis Representation and Decision Boundary
   • Sigmoid function and interpreting logistic regression output.
   • Concept of decision boundary in logistic regression.
3. Cost Function in Logistic Regression
   • Analyzing the cost function.
   • Convex optimization in classification.

Lecture 3: Basics of Neural Networks (2 hours)

1. Introduction to Neural Networks
   • Definition and motivation for using neural networks.
   • Overview of non-linear classification.
2. Perceptrons as Linear Classifiers
   • Introduction to the perceptron model.
   • Understanding how perceptrons work in linear classification.
3. From Linear to Non-Linear Boundaries
   • Limitations of linear classifiers like perceptrons.
   • The need for non-linear classification in complex problems.
4. Introduction to Neural Network Architecture
   • Basic components: neurons, weights, biases.
   • Understanding neuron layers and structure.
5. Learning with Neural Networks
   • Concept of feedforward and backpropagation.
   • Overview of the training process using examples.
6. Intuition on Backpropagation
   • Detailed explanation of backpropagation.
   • The significance of training neural networks.

Lecture 4: Advanced Concepts and Convolutional Neural Networks (2 hours)

1. Deep Dive into Backpropagation
   • In-depth analysis of the backpropagation process.
   • Computational graphs and gradient computation.
2. Understanding Convolutional Neural Networks (CNNs)
   • Distinction between Fully-Connected Networks (FCNs) and CNNs.
   • The architecture and operation of CNNs.
3. Components of CNNs
   • Convolutional layers, activation functions, and pooling.
   • Importance of ReLU, softmax functions, and dropout layers.
4. Training and Optimizing CNNs
   • Forward and backward propagation in CNNs.
   • Strategies to address vanishing/exploding gradients.
5. Application of CNNs in Image Recognition
   • How CNNs achieve feature extraction in image classification.
   • Practical examples of CNNs in image recognition.

Lecture 5: Case study for the detection of melanoma (1 hour)

1. Dealing with medical data
   • Bias in the data
2. Challenges of Unbalanced Datasets
   • Understanding the nature of unbalanced datasets in medical imaging.
   • Introduction to Precision, Recall, F1 Score, AUC.
3. Combining different data sources
   • Demographic information
   • Clinical history
   • Genomics
4. Importance of Accurate and Clinically Relevant Models
   • Relevant metrics in medical imaging: Sensitivity and Specificity

Schedule:

1. May 27, 2024: 14:00-17:00 (Lectures 1 and 2)
2. May 28, 2024: 14:00-16:00 (Lecture 3)
3. May 30, 2024: 14:00-16:00 (Lecture 4)
4. May 31, 2024: 14:00-15:00 (Lecture 5)