Investigating a Feature Unlearning Bias Mitigation Technique for Cancer-type Bias in AutoPet Dataset

Integrating datasets from different cancer types can improve diagnostic accuracy, as deep learning models tend to generalise better with more data. However, this benefit is often limited by performance variance caused by biases, such as under- or over-representation of certain diseases. In this work, we propose a cancer-type-invariant model capable of segmenting tumours from both lymphoma and lung cancer, irrespective of their frequency or representation bias. We frame the problem as a transfer learning task; we introduce a discriminator dedicated to learning bias-group specific features and a confusion loss that preserve generic features while unlearning the domain-specific ones.

Duc Thang Hoang, Quentin Ferre, Elsa Schalck, Olivier Humbert, Rosana El Jurdi, Investigating a Feature Unlearning Bias Mitigation Technique for Cancer-type Bias in AutoPet Dataset, In Proc. of the 30th Colloque Francophone de Traitement du Signal et des Images, Août 2025.

Click here to access the poster.