On 19th of November, 2026, at 15.00 in person, at the Faculty of Science and Technology, University of Latvia, (Jelgavas iela 3) Room 401, as well as remotely, the UL Promotional Council of Physics and Astronomy, and Material Science will hold an open meeting, during which Nicola Rizzieri will defend his doctoral thesis (Ph.D) in natural science. Topic: "Development of a clinician-driven no-code artificial intelligence workflow for retinal image analysis".

Scientific supervisor: Dr. habil. phys. Māris Ozoliņš.

Abstract:

Artificial intelligence (AI) and computer vision are increasingly transforming ophthalmic imaging by enabling automated analysis of retinal fundus photographs for diagnostic and screening purposes. However, the development of high-performance AI systems in medicine still frequently relies on advanced programming expertise, potentially limiting direct clinician involvement in computational workflow design. This doctoral thesis investigated the feasibility, robustness, and translational potential of a clinician-driven no-code AI framework for retinal image analysis using accessible computer vision platforms based on the You Only Look Once (YOLO) architecture. The experimental workflow progressively evolved from established to less explored ophthalmic applications, including diabetic retinopathy (DR) lesion segmentation, optic disc (OD) localisation, binary myopia classification from full-field fundus photographs, and anatomically guided classification using an OD-centred region of interest (ROI). The models’ performance was evaluated through internal validation, independent testing, statistical comparisons, and benchmarking against expert ophthalmologists. In DR analysis, OD segmentation achieved an average precision (AP) of 0.982, whereas microaneurysms segmentation remained substantially more challenging (AP = 0.265). For myopia classification, OD-centred analysis achieved an area under the curve (AUC) of 0.956, with sensitivity of 1.000 and specificity of 0.933, exceeding the discrimination obtained using full-field fundus photographs. In the independent full-field test set, YOLOv11-nano achieved an AUC of 0.889 and YOLOv8-medium an AUC of 0.806, compared with 0.832 for the clinical consensus. No statistically significant difference was detected between YOLOv8-m and the clinical consensus in paired classification outcomes (exact McNemar test, p = 0.302), whereas the difference between YOLOv11-n and the consensus approached but did not reach statistical significance (p = 0.065). Overall, the findings support the feasibility of clinician-driven no-code AI workflows for retinal image analysis and demonstrate the potential value of anatomically guided, clinician-informed model development. The progressive application of the same workflow from established retinal image-analysis tasks to the comparatively underexplored classification of refractive status provides a practical framework for clinician participation in the development of ophthalmic AI. Further prospective, multicentre studies with external validation, age-balanced cohorts, and calibration assessment are required to establish generalisability and clinical utility.

Reviewers:

  1.  Dr. phys. Sergejs Fomins, University of Latvia Institute of Solid State Physics;
  2. Dr. sc. comp. Jānis Bičevskis, University of Latvia;
  3. Dr. Daniel Ledwon, Silesian University of Technology, Poland.

 The doctoral thesis is available at the LU Library, Raiņa bulvāris 19.

Participation in the online session is by prior application, writing to sintija.silina@lu.lv by 13th of November.