Medical Imaging Data Pipelines for Clinical AI

Lectures

• From Pixels to Pathology: What Medical Images Really Represent 

  • T. Calimeri and A. Marzullo

  • This session - split into two parts - sets the stage by grounding both technical and clinical participants in the nature and complexity of medical imaging data. It explores the biology behind the images, how clinicians interpret them, and what technical details underlie their acquisition and structure. By exposing both sides to each other's implicit knowledge, this session breaks down silos and builds a shared understanding of what’s in an image, and what is not.

  • Key Activities:

    • walk through real anonymized case images (CT, pathology, ultrasound, etc.).

    • explain raw data characteristics (DICOM, bit depth, voxel resolution).

    • Group analysis of what’s “easy” or “hard” to detect — and why.

  • Outcome:

    • Shared vocabulary.

    • Realistic appreciation of the visual + biological complexity of images.

• The Data Dilemma: From Hospital Systems to AI Pipelines 

  • A. Marzullo and T. Calimeri

  • Medical data is often seen as abundant, but in reality it’s fragmented, biased, and locked inside complex systems. This session exposes the entire data lifecycle — from clinical acquisition to machine-learning readiness. Participants learn the ethical, technical, and organizational barriers to accessing usable data and the real-world consequences of poor data practices on model performance and patient outcomes.

  • Key Activities:

    • Overview of PACS, DICOM, relevant image characteristics, bias and accountability, data privacy and protection, AI Act and Ethics guidelines for trustworthy AI.

    • Group discussion on what makes data clean, relevant, and representative.

    • Group discussion on possible biases affecting model performances

  • Outcome:

    • Engineers understand clinical IT constraints and regulation.

    • Clinicians see how poor data curation harms models.

  • Foundation for practical collaboration during development.

• Annotation Is Modeling: Designing Data That Teaches

  • A. Marzullo and T. Calimeri

  • Before you train a model, you must define what it should learn — and that happens through annotation. This session unpacks the deceptively difficult process of labeling medical images. Participants explore real examples of ambiguity, disagreement, and downstream effects of label quality. They also learn about smarter annotation strategies and tools, and discuss what “ground truth” even means in medicine.

  • Key Activities:

    • Discuss ambiguous cases and inter-rater variability.

    • Introduce concepts like active learning, semi-supervised labeling, and expert disagreement.

    • Teams critique annotation strategies and simulate decision trade-offs (speed, expertise, structure).

  • Outcome:

    • Clinicians appreciate the value of structured annotation and reproducibility.

    • Engineers recognize labeling as a strategic modeling step, not a pre-modeling chore.

    • Teams align on the idea that annotation design = model design.

• Practical session: design a computer vision-based application for medical image analysis

  • A. Marzullo, S. Moccia, S. Tomassini

  • Shift both clinicians and engineers away from hype and toward meaningful, solvable problems. Goal of this hands-on is to design (and possibly implement parts of) a computer vision-based application for medical image analysis. We start communicating clearly about goals, constraints, and utility.

  • What happens:

    • We will brainstorm “typical” real-world challenges in the healthcare sector..

    • Small mixed groups translate the clinical scenario into ML questions: Is this a classification problem? A detection problem? A risk stratification tool?

    • Engineers explain what’s feasible; doctors judge if it’s useful.

  • Key takeaways:

    • Not everything needs AI.

    • A good problem definition is 50% of the work.

    • Learn each other’s vocabulary and assumptions.

Speakers

• TERESA CALIMERI, MD PhD

  • Senior Staff Physician Lymphoma Unit, IRCCS San Raffale Hospital, Milan, Italy

  • Teresa Calimeri works in the Lymphoma Unit of the IRCCS San Raffaele Hospital, Milan, Italy as a senior staff physician and she has been recently identified as the clinical coordinator of the Disease Unit Lymphoma. Dr. Calimeri’s personal research focus includes the molecular characterization of primary and secondary CNS lymphomas along with the applications of circulating tumor DNA (ctDNA) as a biomarker of diagnosis, treatment response and disease prognosis. Dr. Calimeri has also recently promoted and launched a collaboration with biomedical and informatic engineers with the aim to apply radiomic and deep learning approaches for the studies of lymphoid malignancies. Moreover, she is deeply involved in the clinical management of indolent and aggressive lymphomas. She spent part of her PhD in Molecular Oncology at The Jerome Lipper Multiple Myeloma Center at Dana-Farber Cancer Institute & Harvard Medical School in Boston. In 2017, she also completed a II Level Master course in the diagnosis and therapy of Lymphomas at the University of Udine sponsored by the Fondazione Italiana Linfomi. She is member of both European and International PCNSL Collaborative Group (EPCG and IPCG). She is actively involved in the projects of the Fondazione Italiana Linfomi (FIL) and International Extranodal Lymphoma Study Group (IELSG).

• ALDO MARZULLO, double PhD 

  • Humanitas Research Hospital, Milan, Italy

  • Aldo Marzullo is a postdoctoral researcher with interest in Deep Learning and Medical Imaging. He earned his Ph.D. from the University of Calabria and Université Claude Bernard Lyon 1, with a thesis entitled "Deep Learning and Graph Theory for Brain Connectivity Analysis in Multiple Sclerosis". His current work focuses on anomaly detection in medical imaging, model explainability, and related areas at the intersection of artificial intelligence and healthcare.

• SARA MOCCIA, PhD

  • University “G. D’Annunzio”, Chieti, Italy

  • Sara Moccia earned her master’s degree (cum laude) in Biomedical Engineering from Politecnico di Milano in 2014 and her European PhD (cum laude) in Bioengineering from the Italian Institute of Technology in 2018. During her doctorate she was a visiting researcher at the German Cancer Research Center (DKFZ). She worked as a post‑doctoral researcher at Marche Polytechnic University until 2021, when she became a tenure‑track researcher (RTD‑a) at Scuola Superiore Sant’Anna, a position she held until 2024. She is currently an Associate Professor of Bioengineering at the “G. d’Annunzio” University of Chieti-Pescara. Her research aims to unlock the potential of deep learning for analysing a broad spectrum of medical images, providing support to physicians during clinical and surgical procedures. Her work has received several honours, including the “Gruppo Nazionale di Bioingegneria & Patron” award for her PhD thesis. In 2021 she was awarded a L’Oréal Italia for Women and Science Fellowship in collaboration with the Italian National Commission for UNESCO. She has participated in numerous national and international research projects as a unit leader and is currently the principal investigator of a FISA (Italian Fund for Applied Sciences) project. To date she has overseen roughly €1.5 million in research funding. She is the author of more than 60 articles in international journals. She serves as an Associate Editor for IEEE Transactions on Medical Robotics and Bionics and Medical and Biological Engineering and Computing.

• SELENE TOMASSINI, PhD

  • University of Trento, Italy 

  • Selene Tomassini earned her PhD (2023) in Information Engineering, curriculum Biomedical, Electronic and Telecommunication Engineering, from Università Politecnica delle Marche, defending a thesis entitled “On-cloud decision-support algorithms driven by deep learning in 3D radiological imaging diagnostics”. During her doctoral studies, she focused on the development of AI-based decision-support systems in medical imaging, with particular emphasis on oncology and neurodegeneration. She is currently an Assistant Professor at the Department of Information Engineering and Computer Science, University of Trento, and she teaches “Computer Science Applied to Radiological Sciences” and “AI in Radiodiagnostics and Radiotherapy”. Her research aims to design and develop deep learning models for the processing and analysis of medical images, supporting clinical workflows and diagnostic accuracy, particularly in radiology. She has authored more than 30 articles in international peer-reviewed journals and conferences, and serves as an Associate Editor for Scientific Reports (Nature Portfolio). Her work spans several interdisciplinary collaborations and contributes to advancing the role of AI in healthcare. She is also involved in national and international research initiatives focused on AI for medical imaging.