Publication
Bouzon, P. H. G., da Rocha, W. F., de Souza, L. A., Pacheco, A. G. C. PRISM: a clinically interpretable stepwise framework for multimodal skin cancer diagnosis. Scientific Reports, vol. 16, 2026.
All projects
Journal paper · Scientific Reports
PRISM
Probabilistic Reasoning Interpretable Stepwise Model
The problem
Most computer-assisted diagnostic systems are opaque. A clinician sees a label and a confidence score, with no way to tell how a particular clinical fact — the patient's age, whether the lesion itches, where it is on the body — moved the decision.
That matters practically: a system that cannot show its reasoning cannot be audited, and a system that is confidently wrong is more dangerous than one that says it does not know.
The approach
-
Image prediction as a prior
The vision backbone produces an initial distribution over lesion classes. Everything after that is evidence updating rather than a single opaque forward pass.
-
Stepwise Bayesian updating over clinical features
Each clinical attribute is incorporated one at a time, so the framework can be evaluated with incrementally available metadata. The intermediate states are the explanation — you can read off which feature moved the posterior and by how much.
-
Stepwise calibration
Sequential Bayesian updating compounds overconfidence. The calibration protocol scales with the volume of evidence seen so far, keeping the confidence estimate statistically reliable at every intermediate step, not only at the end.
-
Validated across backbones and datasets
Four competitive vision backbones on three datasets, compared against state-of-the-art attention-based multimodal methods.
Results
PAD-UFES-20, peak result across the evaluated backbones.
77.2 ± 3.6 %
Balanced accuracy
4
Vision backbones
3
Datasets
≥ attention-based SOTA
Comparison
Qualitative case studies show a reasoning process consistent with clinical logic.
Publication & code
Code
Public artefacts related to this work:
Related work
Other parts of the same research line.