Our Connected Imaging business area provides solutions for secure information sharing and collaboration across healthcare domains and a complete image management solution for securely capturing, storing and sharing medical images. The Mobile Telemedicine business unit provides Dermicus, a digital teledermatology solution for fast and secure diagnosis of skin cancer and wounds.

In the development of AI models for skin lesion classification the learning process of the models require large amounts of data. However, skin cancer datasets are often highly imbalanced, with large numbers of benign cases such as nevi and seborrheic keratoses, but few examples of rarer and more serious diseases like squamous cell carcinoma (SCC). This imbalance limits the model’s ability to learn accurate representations for uncommon and difficult classes.

In previous attempts to handle this imbalance augmentation, weighted sampling and focal loss has provided some performance improvements, but performance for SCC continues to be significantly lower compared to other classes on test sets. Curriculum learning offers an alternative approach, where the model is trained in stages, for example, starting with a balanced or easier subset before progressing to the full dataset. This gradual learning process can help the model develop stronger representations for minority classes and ultimately improve classification performance on rare skin cancers.

Master thesis purpose and objectives:

This project aims to investigate curriculum learning as a strategy to mitigate class imbalance in multi-class skin cancer classification. By structuring the training process to gradually increase task difficulty or data imbalance, the aim is to help the network form robust, generalizable features for minority classes before being exposed to a large amount of majority-class examples, ultimately improving model performance.

The objectives of this project are:

• Research and evaluate existing curriculum learning approaches for imbalanced medical image classification.
• Develop and implement proof-of-concept curriculum learning strategies for a multi-class skin cancer model.
• Compare curriculum learning with traditional imbalance-handling methods such as weighted sampling, augmentation, and focal loss.
• Analyze and quantify the impact on overall and class-specific performance, focusing on improvements in minority classes.

Our tech stack:

• This project will be conducted in Python using PyTorch as the main framework.

What we offer:

• Access to data and development environment.
• Mentor and Omda employees' domain knowledge.

Competence requirements:

• Studies engineering physics/mathematics, biomedical engineering, computer science, or similar
• Experience in programming in Python
• Experience with machine-learning and AI projects
• Being motivated, creative, focused, and has problem-solving skills.
• Curiosity and interest in the MedTech area.

Looking forward to meeting you!