| Nome do aluno | Matheus Augusto Oliveira dos Santos |
|---|---|
| Título do trabalho | Prototypical Linear Mapping from Vision Foundation Models for Histopathology Image Retrieval |
| Resumo do trabalho | The rapid expansion of high-resolution histological slide digitization (Whole-Slide Imaging, WSI) has transformed digital histopathology, generating vast repositories of gigapixel-scale slides. This advancement enables data-driven diagnostics but also introduces significant challenges related to storage, annotation, and efficient image retrieval. Content-Based Medical Image Retrieval (CBMIR) systems have emerged as a promising solution to address these challenges, allowing the retrieval of biopsy samples that are visually and/or semantically similar based on image content rather than textual metadata. Although deep learning and contrastive representation learning have driven substantial progress in CBMIR, current models still rely on large labeled datasets and exhibit limited robustness to domain shifts caused by variations in staining and morphology. Recent advances in self-supervised and Transformer-based architectures have given rise to visual foundation models (FMs)—such as DINO, iBOT, and DINOv2—capable of learning transferable visual representations from large volumes of unlabeled data. Building upon these generic models, domain-specific foundation models for histopathology—such as UNI, Virchow, and Phikon—have been trained directly on massive collections of histological slides. These models encode complex morphological and staining patterns across different tissue types. However, their potential for medical image retrieval tasks remains underexplored, and their embeddings are not specifically optimized to capture the subtle morphological continuities required in CBMIR applications. Within this context, this dissertation investigates how foundation models can be linearly adapted and calibrated for the problem of histopathological image retrieval. We propose a lightweight transfer learning framework based on few-shot and prototype-based adaptation, which refines pretrained embeddings into a compact, retrieval-oriented latent space. By imposing a prototype-centered metric alignment under minimal supervision, the method promotes greater intra-class compactness and inter-class separability while preserving the global semantic structure inherited from the foundation model. Extensive experiments conducted on three biomedical datasets—renal glomerulus biopsies, ovarian cancer histology, and dermatoscopic skin lesions—demonstrate that the proposed adaptation consistently improves retrieval accuracy by more than 10 percentage points in mean Average Precision at K (mAP@K), compared to the pretrained models without adaptation. Permutation-based statistical tests confirm the significance of these improvements. Additional qualitative analyses reveal that the adapted embeddings form more coherent and linearly separable clusters, better reflecting diagnostic morphology. Overall, this work bridges the gap between general-purpose visual foundation models and medical image retrieval in domain-specific contexts. The proposed approach provides a scalable, interpretable, and computationally efficient pathway for deploying foundation model–based CBMIR systems in digital pathology, supporting diagnostic reasoning and clinical decision-making with minimal reliance on labeled data. |
| Orientador | Luciano Rebouças de Oliveira |
| Membro Titular Externo (com afiliação) | Jefferson Fontinele da Silva (UFMA) |
| Link para o curriculum lattes |
http://lattes.cnpq.br/ |
| Membro Titular Interno ou Titular Externo 2 (com afiliação) | Angelo Amâncio Duarte (UEFS) |
| Link para o curriculum lattes |
http://lattes.cnpq.br/ |
| Membro Suplente Externo (com afiliação) | Marcelo Mendonça dos Santos (IFBA) |
| Link para o curriculum lattes |
http://lattes.cnpq.br/ |
| Membro Suplente Interno ou Suplente Externo 2 (com afiliação) | Kalyf Abdalla Buzar Lima (IFBA) |
| Link para o curriculum lattes |
http://lattes.cnpq.br/ |
| Data da defesa | 18 Dec, 2025 |
| Horário da defesa | 2:00 PM |
| Quais os principais impactos deste trabalho (social, tecnológico, científico, ambiental)? | Científico: Tratamento linear de embeddings para sistema de recuperação de imagens de forma semi-supervisionada |
Data da Defesa:
18/12/2025 - 14:00
Tipo de Defesa:
Defesa de Mestrado