Name: PEDRO HENRIQUE GALDINO BOUZON
Publication date: 07/08/2026
Examining board:
Name |
Role |
|---|---|
| ANDRE GEORGHTON CARDOSO PACHECO | Presidente |
| GIOVANNI VENTORIM COMARELA | Examinador Interno |
| LUIS ANTONIO DE SOUZA JUNIOR | Coorientador |
| THIAGO MEIRELES PAIXÃO | Examinador Externo |
Summary: Skin cancer remains the most prevalent malignancy worldwide, posing a significant challenge
to global healthcare systems. While multimodal computer-aided diagnosis (CAD) systems
– which integrate clinical metadata with dermatological imaging – have shown promise
in improving diagnostic accuracy, they often face two important limitations: a lack of
resilience to incomplete patient records and a lack of transparency in their decision-making
processes. This work introduces two methods designed to address these challenges. First,
we present MetaBlock-SE, an architecture designed to handle missing or incomplete
metadata. By replacing traditional one-hot encoding with semantic sentence embeddings,
this method captures dense relational features between clinical attributes. Evaluated on
the PAD-UFES-20 dataset and a significantly larger extended version, MetaBlock-SE
demonstrated superior robustness compared to original architectures, achieving balanced
accuracies of up to 69.2 ± 2.8% while remaining computationally efficient for deployment
in resource-constrained environments. Building on this foundation of data resilience, we
also propose PRISM (Probabilistic Reasoning Interpretable Stepwise Model). PRISM
shifts the focus toward clinical transparency by employing a Bayesian framework to model
the diagnostic process as a series of incremental evidence updates. This approach allows
clinicians to evaluate the specific contribution of each clinical feature to the final prediction.
To ensure statistical reliability during this sequential process, we implement a Stepwise
Calibration protocol that mitigates compounding overconfidence. Validated across multiple
datasets and vision backbones, PRISM achieved a peak balanced accuracy of 77.1 ± 3.7%,
outperforming state-of-the-art attention-based methods. Collectively, these contributions
represent a step towards more robust and interpretable multimodal skin lesion classification.
By reconciling computational efficiency with clinical logic, this research offers a trustworthy
framework for real-world clinical decision support.
