Filtern
Volltext vorhanden
- ja (26)
Dokumenttyp
Schlagworte
- artificial intelligence (26) (entfernen)
Institut
- Institut für Informatik (6)
- Betriebswirtschaftliches Institut (5)
- Klinik und Poliklinik für Mund-, Kiefer- und Plastische Gesichtschirurgie (5)
- Institut für Geographie und Geologie (3)
- Klinik und Poliklinik für Dermatologie, Venerologie und Allergologie (3)
- Medizinische Klinik und Poliklinik II (2)
- Institut für Sportwissenschaft (1)
- Institut für diagnostische und interventionelle Radiologie (Institut für Röntgendiagnostik) (1)
- Klinik und Poliklinik für Anästhesiologie (ab 2004) (1)
- Klinik und Poliklinik für Nuklearmedizin (1)
- Klinik und Poliklinik für Unfall-, Hand-, Plastische und Wiederherstellungschirurgie (Chirurgische Klinik II) (1)
- Poliklinik für Kieferorthopädie (1)
- Poliklinik für Zahnärztliche Prothetik (1)
Sonstige beteiligte Institutionen
EU-Projektnummer / Contract (GA) number
- 701983 (1)
Even as medical data sets become more publicly accessible, most are restricted to specific medical conditions. Thus, data collection for machine learning approaches remains challenging, and synthetic data augmentation, such as generative adversarial networks (GAN), may overcome this hurdle. In the present quality control study, deep convolutional GAN (DCGAN)-based human brain magnetic resonance (MR) images were validated by blinded radiologists. In total, 96 T1-weighted brain images from 30 healthy individuals and 33 patients with cerebrovascular accident were included. A training data set was generated from the T1-weighted images and DCGAN was applied to generate additional artificial brain images. The likelihood that images were DCGAN-created versus acquired was evaluated by 5 radiologists (2 neuroradiologists [NRs], vs 3 non-neuroradiologists [NNRs]) in a binary fashion to identify real vs created images. Images were selected randomly from the data set (variation of created images, 40%-60%). None of the investigated images was rated as unknown. Of the created images, the NRs rated 45% and 71% as real magnetic resonance imaging images (NNRs, 24%, 40%, and 44%). In contradistinction, 44% and 70% of the real images were rated as generated images by NRs (NNRs, 10%, 17%, and 27%). The accuracy for the NRs was 0.55 and 0.30 (NNRs, 0.83, 0.72, and 0.64). DCGAN-created brain MR images are similar enough to acquired MR images so as to be indistinguishable in some cases. Such an artificial intelligence algorithm may contribute to synthetic data augmentation for "data-hungry" technologies, such as supervised machine learning approaches, in various clinical applications.
Background: Natural language processing (NLP) is a powerful tool supporting the generation of Real-World Evidence (RWE). There is no NLP system that enables the extensive querying of parameters specific to multiple myeloma (MM) out of unstructured medical reports. We therefore created a MM-specific ontology to accelerate the information extraction (IE) out of unstructured text. Methods: Our MM ontology consists of extensive MM-specific and hierarchically structured attributes and values. We implemented “A Rule-based Information Extraction System” (ARIES) that uses this ontology. We evaluated ARIES on 200 randomly selected medical reports of patients diagnosed with MM. Results: Our system achieved a high F1-Score of 0.92 on the evaluation dataset with a precision of 0.87 and recall of 0.98. Conclusions: Our rule-based IE system enables the comprehensive querying of medical reports. The IE accelerates the extraction of data and enables clinicians to faster generate RWE on hematological issues. RWE helps clinicians to make decisions in an evidence-based manner. Our tool easily accelerates the integration of research evidence into everyday clinical practice.
Background
Several recent publications have demonstrated the use of convolutional neural networks to classify images of melanoma at par with board-certified dermatologists. However, the non-availability of a public human benchmark restricts the comparability of the performance of these algorithms and thereby the technical progress in this field.
Methods
An electronic questionnaire was sent to dermatologists at 12 German university hospitals. Each questionnaire comprised 100 dermoscopic and 100 clinical images (80 nevi images and 20 biopsy-verified melanoma images, each), all open-source. The questionnaire recorded factors such as the years of experience in dermatology, performed skin checks, age, sex and the rank within the university hospital or the status as resident physician. For each image, the dermatologists were asked to provide a management decision (treat/biopsy lesion or reassure the patient). Main outcome measures were sensitivity, specificity and the receiver operating characteristics (ROC).
Results
Total 157 dermatologists assessed all 100 dermoscopic images with an overall sensitivity of 74.1%, specificity of 60.0% and an ROC of 0.67 (range = 0.538–0.769); 145 dermatologists assessed all 100 clinical images with an overall sensitivity of 89.4%, specificity of 64.4% and an ROC of 0.769 (range = 0.613–0.9). Results between test-sets were significantly different (P < 0.05) confirming the need for a standardised benchmark.
Conclusions
We present the first public melanoma classification benchmark for both non-dermoscopic and dermoscopic images for comparing artificial intelligence algorithms with diagnostic performance of 145 or 157 dermatologists. Melanoma Classification Benchmark should be considered as a reference standard for white-skinned Western populations in the field of binary algorithmic melanoma classification.
Background
Melanoma is the most dangerous type of skin cancer but is curable if detected early. Recent publications demonstrated that artificial intelligence is capable in classifying images of benign nevi and melanoma with dermatologist-level precision. However, a statistically significant improvement compared with dermatologist classification has not been reported to date.
Methods
For this comparative study, 4204 biopsy-proven images of melanoma and nevi (1:1) were used for the training of a convolutional neural network (CNN). New techniques of deep learning were integrated. For the experiment, an additional 804 biopsy-proven dermoscopic images of melanoma and nevi (1:1) were randomly presented to dermatologists of nine German university hospitals, who evaluated the quality of each image and stated their recommended treatment (19,296 recommendations in total). Three McNemar's tests comparing the results of the CNN's test runs in terms of sensitivity, specificity and overall correctness were predefined as the main outcomes.
Findings
The respective sensitivity and specificity of lesion classification by the dermatologists were 67.2% (95% confidence interval [CI]: 62.6%–71.7%) and 62.2% (95% CI: 57.6%–66.9%). In comparison, the trained CNN achieved a higher sensitivity of 82.3% (95% CI: 78.3%–85.7%) and a higher specificity of 77.9% (95% CI: 73.8%–81.8%). The three McNemar's tests in 2 × 2 tables all reached a significance level of p < 0.001. This significance level was sustained for both subgroups.
Interpretation
For the first time, automated dermoscopic melanoma image classification was shown to be significantly superior to both junior and board-certified dermatologists (p < 0.001).
The rating of perceived exertion (RPE) is a subjective load marker and may assist in individualizing training prescription, particularly by adjusting running intensity. Unfortunately, RPE has shortcomings (e.g., underreporting) and cannot be monitored continuously and automatically throughout a training sessions. In this pilot study, we aimed to predict two classes of RPE (≤15 “Somewhat hard to hard” on Borg’s 6–20 scale vs. RPE >15 in runners by analyzing data recorded by a commercially-available smartwatch with machine learning algorithms. Twelve trained and untrained runners performed long-continuous runs at a constant self-selected pace to volitional exhaustion. Untrained runners reported their RPE each kilometer, whereas trained runners reported every five kilometers. The kinetics of heart rate, step cadence, and running velocity were recorded continuously ( 1 Hz ) with a commercially-available smartwatch (Polar V800). We trained different machine learning algorithms to estimate the two classes of RPE based on the time series sensor data derived from the smartwatch. Predictions were analyzed in different settings: accuracy overall and per runner type; i.e., accuracy for trained and untrained runners independently. We achieved top accuracies of 84.8 % for the whole dataset, 81.8 % for the trained runners, and 86.1 % for the untrained runners. We predict two classes of RPE with high accuracy using machine learning and smartwatch data. This approach might aid in individualizing training prescriptions.
Hintergrund
Die Fotodokumentation von offenen Frakturen, Wunden, Dekubitalulzera, Tumoren oder Infektionen ist ein wichtiger Bestandteil der digitalen Patientenakte. Bisher ist unklar, welchen Stellenwert diese Fotodokumentation bei der Abrechnungsprüfung durch den Medizinischen Dienst der Krankenkassen (MDK) hat.
Fragestellung
Kann eine Smartphone-basierte Fotodokumentation die Verteidigung von erlösrelevanten Diagnosen und Prozeduren sowie der Verweildauer verbessern?
Material und Methoden
Ausstattung der Mitarbeiter mit digitalen Endgeräten (Smartphone/Tablet) in den Bereichen Notaufnahme, Schockraum, OP, Sprechstunden sowie auf den Stationen. Retrospektive Auswertung der Abrechnungsprüfung im Jahr 2019 und Identifikation aller Fallbesprechungen, in denen die Fotodokumentation eine Erlösveränderung bewirkt hat.
Ergebnisse
Von insgesamt 372 Fallbesprechungen half die Fotodokumentation in 27 Fällen (7,2 %) zur Bestätigung eines Operationen- und Prozedurenschlüssels (OPS) (n = 5; 1,3 %), einer Hauptdiagnose (n = 10; 2,7 %), einer Nebendiagnose (n = 3; 0,8 %) oder der Krankenhausverweildauer (n = 9; 2,4 %). Pro oben genanntem Fall mit Fotodokumentation ergab sich eine durchschnittliche Erlössteigerung von 2119 €. Inklusive Aufwandpauschale für die Verhandlungen wurde somit ein Gesamtbetrag von 65.328 € verteidigt.
Diskussion
Der Einsatz einer Smartphone-basierten Fotodokumentation kann die Qualität der Dokumentation verbessern und Erlöseinbußen bei der Abrechnungsprüfung verhindern. Die Implementierung digitaler Endgeräte mit entsprechender Software ist ein wichtiger Teil des digitalen Strukturwandels in Kliniken.
In Earth observation (EO), large-scale land-surface dynamics are traditionally analyzed by investigating aggregated classes. The increase in data with a very high spatial resolution enables investigations on a fine-grained feature level which can help us to better understand the dynamics of land surfaces by taking object dynamics into account. To extract fine-grained features and objects, the most popular deep-learning model for image analysis is commonly used: the convolutional neural network (CNN). In this review, we provide a comprehensive overview of the impact of deep learning on EO applications by reviewing 429 studies on image segmentation and object detection with CNNs. We extensively examine the spatial distribution of study sites, employed sensors, used datasets and CNN architectures, and give a thorough overview of applications in EO which used CNNs. Our main finding is that CNNs are in an advanced transition phase from computer vision to EO. Upon this, we argue that in the near future, investigations which analyze object dynamics with CNNs will have a significant impact on EO research. With a focus on EO applications in this Part II, we complete the methodological review provided in Part I.
Deep learning (DL) has great influence on large parts of science and increasingly established itself as an adaptive method for new challenges in the field of Earth observation (EO). Nevertheless, the entry barriers for EO researchers are high due to the dense and rapidly developing field mainly driven by advances in computer vision (CV). To lower the barriers for researchers in EO, this review gives an overview of the evolution of DL with a focus on image segmentation and object detection in convolutional neural networks (CNN). The survey starts in 2012, when a CNN set new standards in image recognition, and lasts until late 2019. Thereby, we highlight the connections between the most important CNN architectures and cornerstones coming from CV in order to alleviate the evaluation of modern DL models. Furthermore, we briefly outline the evolution of the most popular DL frameworks and provide a summary of datasets in EO. By discussing well performing DL architectures on these datasets as well as reflecting on advances made in CV and their impact on future research in EO, we narrow the gap between the reviewed, theoretical concepts from CV and practical application in EO.
As part of the Clash of Realities International Conference on the Technology and Theory of Digital Games, the Game Technology Summit is a premium venue to bring together experts from academia and industry to disseminate state-of-the-art research on trending technology topics in digital games. In this first iteration of the Game Technology Summit, we specifically paid attention on how the successes in AI in Natural User Interfaces have been impacting the games industry (industry track) and which scientific, state-of-the-art ideas and approaches are currently pursued (scientific track).
Immersive virtual environments provide users with the opportunity to escape from the real world, but scripted dialogues can disrupt the presence within the world the user is trying to escape within. Both Non-Playable Character (NPC) to Player and NPC to NPC dialogue can be non-natural and the reliance on responding with pre-defined dialogue does not always meet the players emotional expectations or provide responses appropriate to the given context or world states. This paper investigates the application of Artificial Intelligence (AI) and Natural Language Processing to generate dynamic human-like responses within a themed virtual world. Each thematic has been analysed against humangenerated responses for the same seed and demonstrates invariance of rating across a range of model sizes, but shows an effect of theme and the size of the corpus used for fine-tuning the context for the game world.
Modern immersive multimodal technologies enable the learners to completely get immersed in various learning situations in a way that feels like experiencing an authentic learning environment. These environments also allow the collection of multimodal data, which can be used with artificial intelligence to further improve the immersion and learning outcomes. The use of artificial intelligence has been widely explored for the interpretation of multimodal data collected from multiple sensors, thus giving insights to support learners’ performance by providing personalised feedback. In this paper, we present a conceptual approach for creating immersive learning environments, integrated with multi-sensor setup to help learners improve their psychomotor skills in a remote setting.
Today, intelligent systems that offer artificial intelligence capabilities often rely on machine learning. Machine learning describes the capacity of systems to learn from problem-specific training data to automate the process of analytical model building and solve associated tasks. Deep learning is a machine learning concept based on artificial neural networks. For many applications, deep learning models outperform shallow machine learning models and traditional data analysis approaches. In this article, we summarize the fundamentals of machine learning and deep learning to generate a broader understanding of the methodical underpinning of current intelligent systems. In particular, we provide a conceptual distinction between relevant terms and concepts, explain the process of automated analytical model building through machine learning and deep learning, and discuss the challenges that arise when implementing such intelligent systems in the field of electronic markets and networked business. These naturally go beyond technological aspects and highlight issues in human-machine interaction and artificial intelligence servitization.
Background
A basic requirement for artificial intelligence (AI)–based image analysis systems, which are to be integrated into clinical practice, is a high robustness. Minor changes in how those images are acquired, for example, during routine skin cancer screening, should not change the diagnosis of such assistance systems.
Objective
To quantify to what extent minor image perturbations affect the convolutional neural network (CNN)–mediated skin lesion classification and to evaluate three possible solutions for this problem (additional data augmentation, test-time augmentation, anti-aliasing).
Methods
We trained three commonly used CNN architectures to differentiate between dermoscopic melanoma and nevus images. Subsequently, their performance and susceptibility to minor changes (‘brittleness’) was tested on two distinct test sets with multiple images per lesion. For the first set, image changes, such as rotations or zooms, were generated artificially. The second set contained natural changes that stemmed from multiple photographs taken of the same lesions.
Results
All architectures exhibited brittleness on the artificial and natural test set. The three reviewed methods were able to decrease brittleness to varying degrees while still maintaining performance. The observed improvement was greater for the artificial than for the natural test set, where enhancements were minor.
Conclusions
Minor image changes, relatively inconspicuous for humans, can have an effect on the robustness of CNNs differentiating skin lesions. By the methods tested here, this effect can be reduced, but not fully eliminated. Thus, further research to sustain the performance of AI classifiers is needed to facilitate the translation of such systems into the clinic.
The expansion of renewable energies is being driven by the gradual phaseout of fossil fuels in order to reduce greenhouse gas emissions, the steadily increasing demand for energy and, more recently, by geopolitical events. The offshore wind energy sector is on the verge of a massive expansion in Europe, the United Kingdom, China, but also in the USA, South Korea and Vietnam. Accordingly, the largest marine infrastructure projects to date will be carried out in the upcoming decades, with thousands of offshore wind turbines being installed. In order to accompany this process globally and to provide a database for research, development and monitoring, this dissertation presents a deep learning-based approach for object detection that enables the derivation of spatiotemporal developments of offshore wind energy infrastructures from satellite-based radar data of the Sentinel-1 mission.
For training the deep learning models for offshore wind energy infrastructure detection, an approach is presented that makes it possible to synthetically generate remote sensing data and the necessary annotation for the supervised deep learning process. In this synthetic data generation process, expert knowledge about image content and sensor acquisition techniques is made machine-readable. Finally, extensive and highly variable training data sets are generated from this knowledge representation, with which deep learning models can learn to detect objects in real-world satellite data.
The method for the synthetic generation of training data based on expert knowledge offers great potential for deep learning in Earth observation. Applications of deep learning based methods can be developed and tested faster with this procedure. Furthermore, the synthetically generated and thus controllable training data offer the possibility to interpret the learning process of the optimised deep learning models.
The method developed in this dissertation to create synthetic remote sensing training data was finally used to optimise deep learning models for the global detection of offshore wind energy infrastructure. For this purpose, images of the entire global coastline from ESA's Sentinel-1 radar mission were evaluated. The derived data set includes over 9,941 objects, which distinguish offshore wind turbines, transformer stations and offshore wind energy infrastructures under construction from each other. In addition to this spatial detection, a quarterly time series from July 2016 to June 2021 was derived for all objects. This time series reveals the start of construction, the construction phase and the time of completion with subsequent operation for each object.
The derived offshore wind energy infrastructure data set provides the basis for an analysis of the development of the offshore wind energy sector from July 2016 to June 2021. For this analysis, further attributes of the detected offshore wind turbines were derived. The most important of these are the height and installed capacity of a turbine. The turbine height was calculated by a radargrammetric analysis of the previously detected Sentinel-1 signal and then used to statistically model the installed capacity. The results show that in June 2021, 8,885 offshore wind turbines with a total capacity of 40.6 GW were installed worldwide. The largest installed capacities are in the EU (15.2 GW), China (14.1 GW) and the United Kingdom (10.7 GW). From July 2016 to June 2021, China has expanded 13 GW of offshore wind energy infrastructure. The EU has installed 8 GW and the UK 5.8 GW of offshore wind energy infrastructure in the same period. This temporal analysis shows that China was the main driver of the expansion of the offshore wind energy sector in the period under investigation.
The derived data set for the description of the offshore wind energy sector was made publicly available. It is thus freely accessible to all decision-makers and stakeholders involved in the development of offshore wind energy projects. Especially in the scientific context, it serves as a database that enables a wide range of investigations. Research questions regarding offshore wind turbines themselves as well as the influence of the expansion in the coming decades can be investigated. This supports the imminent and urgently needed expansion of offshore wind energy in order to promote sustainable expansion in addition to the expansion targets that have been set.
Novel deep learning (DL) architectures, better data availability, and a significant increase in computing power have enabled scientists to solve problems that were considered unassailable for many years. A case in point is the “protein folding problem“, a 50-year-old grand challenge in biology that was recently solved by the DL-system AlphaFold. Other examples comprise the development of large DL-based language models that, for instance, generate newspaper articles that hardly differ from those written by humans. However, developing unbiased, reliable, and accurate DL models for various practical applications remains a major challenge - and many promising DL projects get stuck in the piloting stage, never to be completed. In light of these observations, this thesis investigates the practical challenges encountered throughout the life cycle of DL projects and proposes solutions to develop and deploy rigorous DL models.
The first part of the thesis is concerned with prototyping DL solutions in different domains. First, we conceptualize guidelines for applied image recognition and showcase their application in a biomedical research project. Next, we illustrate the bottom-up development of a DL backend for an augmented intelligence system in the manufacturing sector. We then turn to the fashion domain and present an artificial curation system for individual fashion outfit recommendations that leverages DL techniques and unstructured data from social media and fashion blogs. After that, we showcase how DL solutions can assist fashion designers in the creative process. Finally, we present our award-winning DL solution for the segmentation of glomeruli in human kidney tissue images that was developed for the Kaggle data science competition HuBMAP - Hacking the Kidney.
The second part continues the development path of the biomedical research project beyond the prototyping stage. Using data from five laboratories, we show that ground truth estimation from multiple human annotators and training of DL model ensembles help to establish objectivity, reliability, and validity in DL-based bioimage analyses.
In the third part, we present deepflash2, a DL solution that addresses the typical challenges encountered during training, evaluation, and application of DL models in bioimaging. The tool facilitates the objective and reliable segmentation of ambiguous bioimages through multi-expert annotations and integrated quality assurance. It is embedded in an easy-to-use graphical user interface and offers best-in-class predictive performance for semantic and instance segmentation under economical usage of computational resources.
Contemporary decision support systems are increasingly relying on artificial intelligence technology such as machine learning algorithms to form intelligent systems. These systems have human-like decision capacity for selected applications based on a decision rationale which cannot be looked-up conveniently and constitutes a black box. As a consequence, acceptance by end-users remains somewhat hesitant. While lacking transparency has been said to hinder trust and enforce aversion towards these systems, studies that connect user trust to transparency and subsequently acceptance are scarce. In response, our research is concerned with the development of a theoretical model that explains end-user acceptance of intelligent systems. We utilize the unified theory of acceptance and use in information technology as well as explanation theory and related theories on initial trust and user trust in information systems. The proposed model is tested in an industrial maintenance workplace scenario using maintenance experts as participants to represent the user group. Results show that acceptance is performance-driven at first sight. However, transparency plays an important indirect role in regulating trust and the perception of performance.
Due to computational advances in the past decades, so-called intelligent systems can learn from increasingly complex data, analyze situations, and support users in their decision-making to address them. However, in practice, the complexity of these intelligent systems renders the user hardly able to comprehend the inherent decision logic of the underlying machine learning model. As a result, the adoption of this technology, especially for high-stake scenarios, is hampered. In this context, explainable artificial intelligence offers numerous starting points for making the inherent logic explainable to people. While research manifests the necessity for incorporating explainable artificial intelligence into intelligent systems, there is still a lack of knowledge about how to socio-technically design these systems to address acceptance barriers among different user groups. In response, we have derived and evaluated a nascent design theory for explainable intelligent systems based on a structured literature review, two qualitative expert studies, a real-world use case application, and quantitative research. Our design theory includes design requirements, design principles, and design features covering the topics of global explainability, local explainability, personalized interface design, as well as psychological/emotional factors.
Künstliche Intelligenz (KI) dringt vermehrt in sensible Bereiche des alltäglichen menschlichen Lebens ein. Es werden nicht mehr nur noch einfache Entscheidungen durch intelligente Systeme getroffen, sondern zunehmend auch komplexe Entscheidungen. So entscheiden z. B. intelligente Systeme, ob Bewerber in ein Unternehmen eingestellt werden sollen oder nicht. Oftmals kann die zugrundeliegende Entscheidungsfindung nur schwer nachvollzogen werden und ungerechtfertigte Entscheidungen können dadurch unerkannt bleiben, weshalb die Implementierung einer solchen KI auch häufig als sogenannte Blackbox bezeichnet wird. Folglich steigt die Bedrohung, durch unfaire und diskriminierende Entscheidungen einer KI benachteiligt behandelt zu werden. Resultieren diese Verzerrungen aus menschlichen Handlungen und Denkmustern spricht man von einer kognitiven Verzerrung oder einem kognitiven Bias. Aufgrund der Neuigkeit dieser Thematik ist jedoch bisher nicht ersichtlich, welche verschiedenen kognitiven Bias innerhalb eines KI-Projektes auftreten können. Ziel dieses Beitrages ist es, anhand einer strukturierten Literaturanalyse, eine gesamtheitliche Darstellung zu ermöglichen. Die gewonnenen Erkenntnisse werden anhand des in der Praxis weit verbreiten Cross-Industry Standard Process for Data Mining (CRISP-DM) Modell aufgearbeitet und klassifiziert. Diese Betrachtung zeigt, dass der menschliche Einfluss auf eine KI in jeder Entwicklungsphase des Modells gegeben ist und es daher wichtig ist „mensch-ähnlichen“ Bias in einer KI explizit zu untersuchen.
Pilot study of a new freely available computer-aided polyp detection system in clinical practice
(2022)
Purpose
Computer-aided polyp detection (CADe) systems for colonoscopy are already presented to increase adenoma detection rate (ADR) in randomized clinical trials. Those commercially available closed systems often do not allow for data collection and algorithm optimization, for example regarding the usage of different endoscopy processors. Here, we present the first clinical experiences of a, for research purposes publicly available, CADe system.
Methods
We developed an end-to-end data acquisition and polyp detection system named EndoMind. Examiners of four centers utilizing four different endoscopy processors used EndoMind during their clinical routine. Detected polyps, ADR, time to first detection of a polyp (TFD), and system usability were evaluated (NCT05006092).
Results
During 41 colonoscopies, EndoMind detected 29 of 29 adenomas in 66 of 66 polyps resulting in an ADR of 41.5%. Median TFD was 130 ms (95%-CI, 80–200 ms) while maintaining a median false positive rate of 2.2% (95%-CI, 1.7–2.8%). The four participating centers rated the system using the System Usability Scale with a median of 96.3 (95%-CI, 70–100).
Conclusion
EndoMind’s ability to acquire data, detect polyps in real-time, and high usability score indicate substantial practical value for research and clinical practice. Still, clinical benefit, measured by ADR, has to be determined in a prospective randomized controlled trial.
Oroantral communication (OAC) is a common complication after tooth extraction of upper molars. Profound preoperative panoramic radiography analysis might potentially help predict OAC following tooth extraction. In this exploratory study, we evaluated n = 300 consecutive cases (100 OAC and 200 controls) and trained five machine learning algorithms (VGG16, InceptionV3, MobileNetV2, EfficientNet, and ResNet50) to predict OAC versus non-OAC (binary classification task) from the input images. Further, four oral and maxillofacial experts evaluated the respective panoramic radiography and determined performance metrics (accuracy, area under the curve (AUC), precision, recall, F1-score, and receiver operating characteristics curve) of all diagnostic approaches. Cohen's kappa was used to evaluate the agreement between expert evaluations. The deep learning algorithms reached high specificity (highest specificity 100% for InceptionV3) but low sensitivity (highest sensitivity 42.86% for MobileNetV2). The AUCs from VGG16, InceptionV3, MobileNetV2, EfficientNet, and ResNet50 were 0.53, 0.60, 0.67, 0.51, and 0.56, respectively. Expert 1–4 reached an AUC of 0.550, 0.629, 0.500, and 0.579, respectively. The specificity of the expert evaluations ranged from 51.74% to 95.02%, whereas sensitivity ranged from 14.14% to 59.60%. Cohen's kappa revealed a poor agreement for the oral and maxillofacial expert evaluations (Cohen's kappa: 0.1285). Overall, present data indicate that OAC cannot be sufficiently predicted from preoperative panoramic radiography. The false-negative rate, i.e., the rate of positive cases (OAC) missed by the deep learning algorithms, ranged from 57.14% to 95.24%. Surgeons should not solely rely on panoramic radiography when evaluating the probability of OAC occurrence. Clinical testing of OAC is warranted after each upper-molar tooth extraction.