A Dempster-Shafer Approach to Trustworthy AI With Application to Fetal Brain MRI Segmentation

Abstract

Deep learning models for medical image segmentation can fail unexpectedly and spectacularly for pathological cases and images acquired at different centers than training images, with labeling errors that violate expert knowledge. Such errors undermine the trustworthiness of deep learning models for medical image segmentation. Mechanisms for detecting and correcting such failures are essential for safely translating this technology into clinics and are likely to be a requirement of future regulations on artificial intelligence (AI). In this work, we propose a trustworthy AI theoretical framework and a practical system that can augment any backbone AI system using a fallback method and a fail-safe mechanism based on Dempster-Shafer theory. Our approach relies on an actionable definition of trustworthy AI. Our method automatically discards the voxel-level labeling predicted by the backbone AI that violate expert knowledge and relies on a fallback for those voxels. We demonstrate the effectiveness of the proposed trustworthy AI approach on the largest reported annotated dataset of fetal MRI consisting of 540 manually annotated fetal brain 3D T2w MRIs from 13 centers. Our trustworthy AI method improves the robustness of four backbone AI models for fetal brain MRIs acquired across various centers and for fetuses with various brain abnormalities. Our code is publicly available here. ✦ INTRODUCTIONA UTOMATIC segmentation of medical images is needed for personalized medicine and to study anatomical development in healthy populations as well as populations with a pathology. Artificial Intelligence (AI) algorithms for medical image segmentation can reach super-human accuracy on average [1] and yet most radiologists do not trust them [2], [3]. This is partly because, for some cases, AI algorithms fail spectacularly with errors that violate expert knowledge about the segmentation task when the AI was applied across imaging protocol and anatomical pathologies [2], [4], [5] (Fig. 1b). This sense of distrust is exacerbated by the current lack of clear fit-for-purpose regulatory requirements for AI-based medical image software [6].The legal framework for the deployment in clinics of AI tools for medical segmentation is likely to soon become more stringent once the European Union has proposed its Artificial Intelligence Act to regulate AI and AI trust is at the core of this proposal [7], [8]. Guidelines for trustworthy

Towards Unlocking the Hidden Potentials of the Data-Centric AI Paradigm in the Modern Era

“…Artificial intelligence (AI) is a mainstream technology with a wide range of promising applications in different sectors such as healthcare, smart cities, chatbots, etc. The representative AI applications in the medical area are disease classification using convolutional neural network (CNN) by leveraging image data [1], fetal brain MRI segmentation to identify brain abnormalities [2], accurate and effective segmentation of medical images for clinical assessment of different diseases [3], personalized healthcare and medical content generation for personalized medication and surgery planning [4], medical question-answer systems [5], and situational awareness for people who are visually impaired or blind in indoor environments [6], to name just a few. The representative AI applications in industry are product quality and design optimization [7], fault detection and failure mode prediction [8], industrial predictive modeling by extracting salient and far features with CNN [9], predictive maintenance [10], detecting defects in products [11], video surveillance [12], and robust continuous-flow manufacturing processes by imputing missing time series data [13], to name just a few.…”