Analysis of AI Diagnostic Performance Discrepancies Across Medical Imaging Modalities

Keywords: Imaging modality, AI performance gap, Architectural limitation, Hybrid diagnosis, Clinical utility

TL;DR: AI diagnostic accuracy in medical imaging varies by modality—ultrasound outperforms CT/MRI—due to data and model architecture mismatch; a hybrid workflow combining X-ray/ultrasound screening and CT/MRI confirmation is proposed for optimal results.

Abstract: Artificial intelligence (AI) shows immense promise in medical imaging, yet its diagnostic performance varies significantly across different modalities. This discrepancy is highlighted by the "ultrasound paradox," where AI achieves superior performance on comparatively lower-quality ultrasound images (AUROC 0.94) while struggling with high-resolution, complex modalities like MRI (reported accuracy as low as 0%). This suggests that performance is not dictated by image quality alone but by a complex interplay between the data’s intrinsic properties and the structural limitations of current AI architectures. This paper provides a deep-dive analysis of this performance gap by systematically reviewing literature on static, high-contrast (CT, MRI) and dynamic, low-contrast (X-ray, ultrasound) modalities. We investigate the root causes, attributing them to a mismatch between the information type provided by a modality (e.g., spatio-temporal data in ultrasound) and the architectural constraints of dominant AI models like Convolutional Neural Networks (CNNs), such as their limited receptive fields and difficulty in processing temporal features. As a practical solution, we propose a multi-stage "hybrid diagnostic workflow" that strategically combines high-sensitivity AI for initial screening (using X-ray/ultrasound) with high-specificity AI for confirmation (using CT/MRI). This approach aims to optimize overall diagnostic accuracy and clinical efficiency. We conclude that the future of medical AI lies not in a single, universal model but in an integrated, collaborative ecosystem that leverages the unique strengths of different modalities and AI architectures to create robust, clinically-relevant solutions.

Submission Number: 72