Thesis Supervision (77 items)

Abstract: Automatic Identification and Data Capture (AIDC) systems are the technological backbone of modern logistics, retail, and manufacturing, enabling traceability, automation, and process monitoring through the rapid acquisition of visual or coded information. Among these technologies, barcodes remain one of the most widespread and cost-effective solutions for product identification. Yet, despite their maturity, barcode and symbol recognition still face major challenges under real-world industrial conditions, where lighting variations, blur, long acquisition distances, or low sensor resolution can drastically reduce readability. Traditional computer-vision algorithms—based on geometric analysis, morphological operators, or the Hough transform—are reliable in controlled settings but fail when imaging conditions deviate from nominal parameters. Conversely, deep learning offers higher flexibility and robustness but requires heavy computation, limiting its use on embedded hardware. Bridging this gap between accuracy and efficiency is crucial for the next generation of intelligent AIDC systems. This thesis presents an in-depth study of benchmarking, optimizing, and deploying efficient deep-learning models tailored to industrial AIDC applications. The work, carried out in collaboration with Datalogic S.p.A., focuses on integrating adaptive neural architectures into constrained, realtime environments. The first part addresses the scarcity of open datasets and reproducible benchmarks in barcode localization. To fill this gap, the BarBeR – Barcode Benchmark Repository was developed: a public framework supported by an open dataset of 8 748 annotated images. BarBeR unifies classical and deep-learning detection methods under consistent evaluation protocols and metrics, enabling fair comparison and reproducibility. Results confirm that while deep models surpass traditional approaches in accuracy, their computational cost remains a critical bottleneck for real-time operation on embedded platforms. To overcome this limitation, the thesis proposes BaFaLo, a lightweight segmentation-based barcode localizer optimized for CPU-class processors. Building upon the Fast-SCNN design paradigm, BaFaLo achieves a balanced trade-off between speed and precision, detecting even small and degraded barcodes in challenging conditions while maintaining real-time speed. Localization, however, is only the first step: decoding remains infeasible when the resolution is too low. To address this, Mosaic-SR was introduced—an adaptive multi-step super-resolution method that selectively allocates computational effort to complex image regions. Guided by uncertainty estimation, Mosaic-SR improves decoding accuracy and latency compared to uniform SR approaches, making high-quality reconstruction feasible on embedded hardware. The final part, conducted during a visiting period at the Integrated Systems Laboratory, ETH Zurich, focuses on model quantization and deployment. This work demonstrates that combining advanced model quantization strategies, such as VDQuant, with cache quantization can reduce the memory footprint of these models by more than 50%, with minimal impact on image quality or stability. These results pave the way for deploying generative architectures on embedded or resource-constrained platforms and for leveraging them in synthetic data generation when real or open datasets are limited. Overall, the thesis demonstrates how efficient and adaptive deep learning can make advanced vision capabilities accessible to real-time AIDC systems. By benchmarking, optimizing, and deploying neural architectures across detection, enhancement, and generative tasks, this work contributes to the evolution of industrial vision—from rigid, rule-based pipelines to flexible, data-driven solutions that operate reliably under real-world constraints.

Abstract: Cone-beam computed tomography (CBCT) is central to contemporary dental and maxillofacial care, yet progress in automated analysis has been constrained by the paucity of large, publicly available voxel-level datasets. This thesis addressed that bottleneck by creating an open, extensible ecosystem, combining datasets, annotation tooling, algorithmic advances, and by demonstrating how those elements interacted cyclically to accelerate research and clinical translation. The Maxillo dataset was the first of its kind, providing 91 densely annotated volumes plus 256 sparsely annotated scans for the annotation of the Inferior Alveolar Canal. The ToothFairy series built on this foundation: the first ToothFairy release provided 443 CBCT volumes (153 with dense 3D annotations); ToothFairy2 expanded to 480 volumes with 42 semantic classes for comprehensive maxillofacial segmentation; with ToothFairy3 it further grew the corpus to 532 volumes and 77 classes. Complementing volumetric CBCTs, the Bits2Bites dataset delivered 200 registered intra-oral scan pairs with multi-label occlusion annotations. All resources were openly released to enable reproducible benchmarking and downstream development. To scale annotations without sacrificing clinical fidelity, I developed semiautomated annotation tools and a rigorous quality-control pipeline that combined model priors with expert refinement. Crucially, dataset creation, tooling, and model development proceeded cyclically: additional data enabled stronger models; stronger models powered faster, higher-quality annotation tools; and improved tools in turn produced larger, better datasets, forming the core intellectual contribution of this work. On this data foundation, I advanced volumetric segmentation methods: modules inspired by transformer architectures that explicitly encoded spatial patch relationships to preserve voxel detail while aggregating long-range context, and adaptations of State-Space Models (Mamba) for efficient, high-accuracy 3D segmentation. Finally, I introduced U-Net Transplant, a model-merging framework that proposed novel techniques to update and specialize clinical models without full retraining, reducing redeployment cost, storage, and privacy exposure. Collectively, this ecosystem delivered the largest open CBCT benchmark for maxillofacial segmentation to date and a matched set of methods and tools that materially improved accuracy, efficiency, and lifecycle management of clinical AI, enabling faster, safer, and more reproducible dental AI research and deployment.

Abstract: Reproducibility policies aim to make medical-imaging research “checkable” at review time, yet many submissions still ship artifacts that are missing, incomplete, or not practically verifiable under peer-review constraints. In a longitudinal analysis of 3,722 MICCAI papers, the fraction linking code increases from 51.8% to 72.5%, but ~13% of linked repositories are inaccessible or empty, and many others lack the concrete information needed to audit claims without executing untrusted code. This thesis introduces PAPER-SNITCH, a reviewer-facing decision support for evidence-based reproducibility screening. PAPER-SNITCH parses conference metadata and PDFs, resolves and sanity-checks linked repositories, and applies a policy-aware checklist aligned with MICCAI-style guidance. Rather than attempting full reproduction, it performs bounded, inspectable checks (e.g., artifact presence, documentation completeness, environment specification, and claim-to-command traceability) and produces an auditable report that links each criterion outcome to concrete evidence excerpts and repository artifacts. Criterion outcomes are aggregated deterministically into interpretable component scores and a global verifiability summary. We evaluate PAPER-SNITCH against human assessment on a sample of MICCAI papers and analyze practical deployment considerations, including per-paper cost, caching, and observability. Finally, we discuss key limitations—most notably incomplete recall under bounded retrieval, susceptibility to strategically written text, and the non-determinism and drift of LLM-backed components—and outline directions for robustness hardening and reviewer-centered validation.

Abstract: Accurate positioning of orthodontic brackets is a critical step in fixed appliance therapy, as placement errors directly affect treatment efficiency, duration, and clinical outcomes. Recent advances in digital orthodontics and intra-oral scanning enable the automation of this process through data-driven methods. This thesis presents a deep learning framework for predicting dental bracket installation points directly from intra-oral scans of patients. The proposed approach operates on three-dimensional dental data and leverages point cloud representations to model tooth geometry. A complete pipeline is developed, including data collection, preprocessing, tooth segmentation, and bracket position prediction. Multiple prediction strategies are investigated, including direct regression and heatmap-based formulations, and their performance is evaluated on curated datasets derived from clinical annotations. To support model deployment in real-world scenarios, a dedicated segmentation model is designed and integrated, providing improved flexibility and control compared to off-the-shelf solutions. Extensive experimental results demonstrate that the proposed methods achieve accurate and consistent bracket placement predictions, approaching clinically acceptable precision. Additionally, the system is packaged and deployed through a web-based API, enabling seamless integration with existing orthodontic software and workflows. This work highlights the feasibility of automated bracket placement from intra-oral scans and provides a scalable foundation for future improvements through additional expert annotations, model refinement, and extended evaluation of segmentation performance. The proposed framework contributes toward reducing manual effort and variability in orthodontic bracket positioning, supporting more efficient and standardized digital orthodontic treatments.

Abstract: Male infertility is a widespread health concern, with testicular ultrasound imaging playing a key role in its assessment. In particular, parenchymal inhomogeneity has been proposed as a biomarker, but its reliable evaluation is challenged by subjective interpretation, artifacts, and the limited availability of annotated datasets. This thesis addresses these challenges by investigating strategies that combine pretraining and synthetic data generation to enhance automated classification of testicular ultrasound images. A ResNet-18 backbone is employed, and both supervised and self-supervised pretraining strategies are evaluated to improve feature representation in data-scarce conditions. To reduce the impact of noisy labels, a heuristic filtering method is proposed, identifying and correcting mislabeled samples. Furthermore, synthetic ultrasound data are generated using diffusion models, followed by a filtering procedure to ensure fidelity and clinical relevance. Experimental results demonstrate that pretraining consistently improves classification performance compared to training from scratch, while synthetic data can effectively support pretraining, partially overcoming data scarcity and privacy limitations. These findings highlight the potential of integrating synthetic data and pretraining strategies to build robust and reliable diagnostic tools, ultimately contributing to the advancement of automated ultrasound analysis for male infertility.

Abstract: The synthesis of missing MRI modalities has emerged as a critical strategy to address incomplete multi-parametric imaging in brain tumor diagnosis and treatment planning. Recent advances in generative models, particularly GANs and diffusion-based approaches, have shown promising results in cross-modality MRI generation, although challenges persist in preserving anatomical fidelity and minimizing synthesis artifacts. Building on the Hybrid Fusion GAN (HF-GAN) framework, several enhancements are introduced to improve synthesis quality and generalization across tumor types. These include the application of z-score normalization, optimization of network components for faster and more stable training, and the extension of the pipeline to support multi-view generation across diverse brain tumor categories such as gliomas, metastases, and meningiomas. The approach emphasizes refinement of 2D slice-based generation to ensure intra-slice coherence and reduce intensity inconsistencies, ultimately facilitating more accurate and robust tumor segmentation in scenarios with missing imaging modalities.