Federated vs local vs central deep learning of tooth segmentation on panoramic radiographs

Artificial Intelligence (AI) has shown great potential to transform dentistry; analyzing a wealth of dental data with AI and using it to support diagnostics, treatment planning and actual treatment has been demonstrated to be feasible across all dental disciplines [1].

Most dental AI employs machine learning, where mathematical models are utilized to identify the inherent structure of a training dataset to allow inference (prediction) on unseen test data. Usually, this involves labeling of data by experts e.g., the classification of an image as showing caries lesions, or detecting the location of a certain pathology on an image etc. [2].

The translation of developed AI models from the research stage into the clinical environment, however, remains slow. Despite a wealth of studies, only a few products have successfully passed regulatory hurdles and entered routine care [3]. The main barrier for this is grounded in the poor generalizability of many AI models. As models are typically trained and tested using data from one center, recorded with one technique, methodology, and represent a single population.

An AI application rarely performs similarly well if applied on data from other centers, gathered using other technical setups, representing different populations, which often differ in age, gender, socio-demographic characteristics, or oral health status. [4] Collaborative efforts (e.g., gathering data from multiple centers) may help to overcome generalizability issues and also allow smaller or less experienced research groups to participate in state-of-the-art AI research. However, such efforts are limited by privacy constraints, which lead to difficulties in exchanging particularly dental data as it is oftentimes hard to de-identify [5].

Federated Learning (FL) is a learning paradigm which enables collaborative, data-driven research between multiple centers through a privacy-by-design approach. It avoids critical exchanges of sensitive data between centers and instead relies on sharing abstract model parameters, which essentially carry the knowledge learned from this data. FL was originally aimed at parallelized training on edge devices and smartphones but has caught considerable attention in healthcare [6], [7], [8], [9] mainly as it may assist to overcome privacy limitations and allow to train generalizable models. However, dental research on FL is still limited [10].

In the present study, we aimed to assess FL for tooth segmentation on panoramic radiographs, a specific (and exemplary) task in dental image analysis. Tooth segmentation involves labeling pixels belonging to each tooth on a panoramic, which allows to identify, classify and relate further findings (e.g., a caries or apical lesion) of an AI-based analysis to a specific tooth. It was further useful for the present study, as tooth segmentation can be relatively easily performed by humans (who label the radiographs before using them for training) and can hence be standardized across centers, reducing the effect of center-specific labeling on the outcomes of FL. We used radiographs from nine international centers and compared FL against Local Learning (LL, involving training on isolated data of each center) and Central Learning (CL, involving data pooling, e.g., under the assumption of data sharing agreements being in place). We also tested models for their generalizability across centers. Our hypothesis was that FL significantly improves the performance and generalizability in comparison with LL (i.e., when CL training is not feasible due to privacy regulations). We further investigated whether specific centers benefited particularly from FL given their specific data distribution.

留言 (0)

沒有登入
gif