Predrag Knežević

Prof. Dr. Predrag Knežević


Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
9 December 2023 in Medicina
Verified
Journal Article
Open access
PDF file
MRMario Rašić
Mario Rašić
MTMario Tropčić
Mario Tropčić
PKPjetra Karlović
Pjetra Karlović
DGDragana Gabric
Dragana Gabric
MSMarko Subašić
Marko Subašić
PKPredrag Knežević
Predrag Knežević
Background and Objectives: The purpose of this study was to develop and evaluate a deep learning model capable of autonomously detecting and segmenting radiolucent lesions in the lower jaw by utilizing You Only Look Once (YOLO) v8. Materials and Methods: This study involved the analysis of 226 lesions present in panoramic radiographs captured between 2013 and 2023 at the Clinical Hospital Dubrava and the School of Dental Medicine, University of Zagreb. Panoramic radiographs included radiolucent lesions such as radicular cysts, ameloblastomas, odontogenic keratocysts (OKC), dentigerous cysts and residual cysts. To enhance the database, we applied techniques such as translation, scaling, rotation, horizontal flipping and mosaic effects. We have employed the deep neural network to tackle our detection and segmentation objectives. Also, to improve our model’s generalization capabilities, we conducted five-fold cross-validation. The assessment of the model’s performance was carried out through metrics like Intersection over Union (IoU), precision, recall and mean average precision (mAP)@50 and mAP@50-95. Results: In the detection task, the precision, recall, mAP@50 and mAP@50-95 scores without augmentation were recorded at 91.8%, 57.1%, 75.8% and 47.3%, while, with augmentation, were 95.2%, 94.4%, 97.5% and 68.7%, respectively. Similarly, in the segmentation task, the precision, recall, mAP@50 and mAP@50-95 values achieved without augmentation were 76%, 75.5%, 75.1% and 48.3%, respectively. Augmentation techniques led to an improvement of these scores to 100%, 94.5%, 96.6% and 72.2%. Conclusions: Our study confirmed that the model developed using the advanced YOLOv8 has the remarkable capability to automatically detect and segment radiolucent lesions in the mandible. With its continual evolution and integration into various medical fields, the deep learning model holds the potential to revolutionize patient care.

22 August 2026 in Children
Verified
Journal Article
Open access
PDF file
ATAntonia Tarle
Antonia Tarle
Marko TarleMarko Tarle
Marko Tarle
MRMarina Raguž
Marina Raguž
SFSanda Huljev Frković
Sanda Huljev Frković
PKPredrag Knežević
Predrag Knežević
Background/Objectives: Cleft palate is developmentally and clinically distinct from cleft lip with or without cleft palate. Prenatal ultrasound reliably identifies clefts involving the lip but performs poorly for the secondary palate, whereas early feeding difficulty is concentrated in palatal clefts. These two observations have not previously been quantified within a single cohort. We aimed to determine whether prenatal detectability and neonatal feeding support requirements follow opposite gradients across cleft phenotypes. Methods: We conducted an ambispective study of a nationally ascertained referral-centre cohort of 328 consecutive children with an orofacial cleft and no clinically recognised syndrome, managed at the national referral centre for cleft surgery in Croatia. The cohort comprises children reaching surgical care and is not a population-based birth registry. Documented prenatal ultrasound diagnosis and feeding method during the first weeks of life were abstracted from medical records; cleft phenotype was confirmed by clinical examination. Proportions were compared with chi-square tests and exact binomial confidence intervals, and factors independently associated with each outcome were identified by multivariable logistic regression. Results: Overall prenatal detection was 76/316 (24.1%; 95% CI 19.4–29.2). Detection differed markedly by phenotype: 38.2% for cleft lip only, 34.0% for cleft lip and palate and 2.0% for cleft palate only (p < 0.001). Nasogastric tube feeding was required by 46/325 children (14.2%; 95% CI 10.6–18.4) and followed the opposite gradient: 0.0%, 14.2% and 24.8%, respectively (p < 0.001). Isolated cleft palate showed the strongest independent negative association with prenatal detection (adjusted OR 0.037; 95% CI 0.009–0.154) and was independently associated with tube feeding (adjusted OR 3.07; 95% CI 1.37–6.89). Of children requiring tube feeding, 43/46 (93.5%; 95% CI 82.1–98.6) had not been detected prenatally. Observed detection rose with year of birth in unadjusted analysis (OR 1.095 per year; 95% CI 1.012–1.185; p = 0.023) but was attenuated and no longer significant after adjustment (OR 1.071; 95% CI 0.991–1.157; p = 0.082); the temporal trend is therefore reported descriptively. An initial refer result at newborn hearing screening was markedly more frequent in children with palatal involvement than in those with cleft lip only (2.8%, 31.6% and 25.5% for cleft lip only, cleft lip and palate and cleft palate only; p < 0.001), contrasting cleft lip only with clefts involving the palate rather than following a monotonic gradient. Conclusions: Prenatal detectability and neonatal nasogastric tube use follow opposite gradients across cleft phenotypes. Children with cleft palate only are simultaneously the least likely to be identified before birth and the most likely to require assisted feeding, and almost all tube-fed infants arrive without a prenatal diagnosis. These findings support, as a practice consideration, routine structured neonatal feeding assessment for every newborn with a palatal cleft irrespective of prenatal findings.