Name: HUMBERTO GIURI CALENTE
Publication date: 12/08/2026
Examining board:
Name |
Role |
|---|---|
| DIEGO ROBERTO COLOMBO DIAS | Examinador Interno |
| GUILHERME PALERMO COELHO | Examinador Externo |
| RENATO ANTONIO KROHLING | Presidente |
Summary: Content-Based Image Retrieval (CBIR) systems are effective tools for supporting clinical
decision-making in dermatology, as they enable the retrieval of historically similar cases.
Despite their usefulness, these systems face the challenge of algorithmic bias, whereby the
underrepresentation of different skin tones can distort results and compromise diagnostic
equity. In this dissertation, a medical CBIR system with integrated bias mitigation
was developed, addressing both feature extraction and result ranking. The proposed
architecture combines Debiasify, a self-distillation-based method for generating bias-
mitigated descriptors, with the TOPSIS multi-criteria decision-making method. This
approach reformulates image retrieval as a decision problem, allowing the final ranked list
to balance visual similarity and demographic equity (skin tone representation). To evaluate
the methodology, experiments were conducted on three skin lesion datasets: PAD-UFES-20,
composed of images captured using smartphones; Fitzpatrick17k, a larger-scale benchmark;
and MRA-MIDAS, which contains prospectively collected images. The results show that
the effects of Debiasify on retrieval performance and fairness depend on the characteristics
of each dataset, whereas TOPSIS consistently reduced representational imbalance in the
retrieved lists, with only small variations in retrieval performance when applied to the
same descriptors.
