20 May 2024
13:00 Doctoral defense IC3 building auditorium
Theme
Robust Skin Lesion Analysis: Evaluation and Debiasing across Distribution Shifts
Student
Alceu Emanuel Bissoto
Advisor / Teacher
Sandra Eliza Fontes de Avila - Co-supervisor: Eduardo Alves do Valle Júnior
Brief summary
Deep learning models are being used in a growing list of real-world applications, including automated diagnostics. Because they are data-driven models, they can replicate biases found in training data. For medical imaging, few centers contribute data for problems of global dimensions, such as skin cancer. The inevitable shift in distribution caused by different populations, hospital procedures, and many other possible sources of bias can lead to catastrophic consequences. In this thesis, we investigate changes in skin injury datasets and models. Skin cancer is an important health problem, being responsible for the majority of cancers in Brazil. Early detection is crucial for a positive prognosis and automated skin cancer detection offers a promising solution, especially for patients facing geographic or economic barriers. Because skin cancer diagnosis by dermatologists makes use of pattern recognition, this task aligns well with machine learning techniques. Skin lesion analysis is a rapidly evolving field, benefiting from a growing volume of data and varied modalities. However, problems such as lack of generalization and excessive reliance on spurious correlations prevent the widespread use of these solutions. Our work contributes to several aspects of this problem, including data annotation, bias assessment, and debiasing. We annotated skin lesion datasets to indicate the presence and location of artifacts, enabling us to assess the robustness of the model and facilitating debiasing. We introduce a new approach to separate training and testing data to evaluate model biases. This method, called "Trap Sets", is designed to reveal a model's dependence on spurious correlations by adjusting bias levels during training and presenting countercorrelations during testing. Trap Sets allow for accurate assessment of learning ``shortcuts'' in skin lesion analysis, something typically limited to simpler, more controlled data sets. Our experiments demonstrate that models rely on these problematic features, a behavior that Trap Sets severely penalize. When addressing bias elimination, we explore training and testing strategies. During training, we leverage our artifact annotations to guide models toward learning features more aligned with medical knowledge. At test time, we focus on identifying and utilizing clinically relevant features for inference. The integration of these strategies proved to be effective, pointing the way to fairer and more effective diagnoses of skin lesions.
Examination Board
Headlines:
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| André Georghton Cardoso Pacheco | INF/UFES |
| Flávia Vasques Bittencourt | FM/UFMG |
| Esther Luna Colombini | IC / UNICAMP |
| Letícia Rittner | FEEC / UNICAMP |
Substitutes:
| Aurea Rossy Soriano Vargas | IC / UNICAMP |
| Ana Gabriela Salvio | FAC |
| Agma Juci Machado Traina | ICMC / USP |