Item 3 — Scientific and/or clinical background, including the intended use and role of the AI approach

The current practice should be explicitly mentioned. Describe the study’s rationale, goals, and anticipated impact. Present a focused summary of the pertinent literature to describe current practice and highlight how the investigation changes or builds on that work. Guide readers to understand the underlying science, assumptions underlying the methodology, and nuances of the study.

Explanation and Elaboration: The introduction should contextualize the work presented in the sections that follow in terms of the significance of the problem or area of inquiry, as well as prior investigations and current state of the art. Understanding the intended use and role of the AI enables the reader to evaluate the reported performance of the AI in an appropriate context, as minimum thresholds for performance vary widely by use case. Description of gaps or limitations in existing knowledge, technology or practice addressed by the manuscript elucidate the advances that have been achieved.

Examples
However, differences in MRI acquisition protocols and scanners affect consistency and reproducibility of brain volumetry … and are a major impediment for the clinical translation of automated tools. To tackle this problem, many data harmonization tools have emerged in recent years…. Such algorithms can either harmonize original scans (eg, DeepHarmony) … or derivatives extracted from the scans (eg, ComBat) … Some of these algorithms have been shown to harmonize patient data affected by a neurodegenerative disease … while preserving disease-related signature. However, such harmonization techniques typically work only for the scanner models they have been trained on, and, in some instances require the same individuals to be scanned with different scanners…. Harmonizing volumetric data from MRI scanners not encountered during initial model training requires additional training with a substantial number of images from these scanners…. This poses a challenge for the deployment of such models for clinical use. [Introduction (10), CC-BY 4.0]
Despite recent advances in the automatic processing of spinal cord MRI…, robust methods for detecting quantitative MRI biomarkers in SCI are lacking. As a result, most studies involve manual identification of these biomarkers…. This task is time-consuming and further hampered by interrater variability, making it impractical in clinical trials…. Furthermore, segmentation of intramedullary SCI lesions on MRI scans poses an extremely challenging task mainly due to the evolving appearance of lesions in different injury phases (eg, acute, subacute, intermediate)…. Surgical implants in postoperative MRI scans might also cause severe image artifacts. Deep learning (DL) can improve diagnosis and prognostication in SCI by automating the lesion annotation process, thereby reducing rater-specific biases and facilitating the analysis of large SCI cohorts across sites…. Indeed, quantitative SCI lesion biomarkers derived from DL-based automatic segmentations have been shown to correlate well with clinical measures of motor impairment…. Despite its numerous potential advantages, DL has not been sufficiently explored in the context of SCI, with no open-source methods existing to date. This gap suggests the need for an automatic biomarker identification method that takes into account the complex pathophysiology of patients with SCI…, generalizes to multiple sites, and remains easily accessible by researchers. [Introduction (11), CC-BY 4.0]
Radiology reports are meticulously crafted and reflect the train of thought and reasoning of radiology experts during interpretation of medical images; they provide detailed descriptions of brain lesions characteristics and anatomic alterations in extensive real-world clinical cases and are organized in a standardized manner. Radiology reports are naturally and seamlessly generated during the clinical diagnostic process, incurring no supplementary production costs. Consequently, they can serve as a high-accuracy, low-cost source of weak annotations for MR images. Previous studies … have accomplished multimodal image-text pretraining through the application of contrastive learning on chest radiographs and corresponding radiology reports. This involved the use of vision and text encoders or pretrained models customized for radiology-specific contexts. In contrast to radiographs, MR images are characterized by multiple sequences, which offer better tissue contrast and can help discern lesions with complex components, leading to more intricate and rich textual features in their reports. An effective diagnostic aid model should provide accurate and stable diagnostic results, offering interpretable indicators to substantiate its predictions as references for radiologists. To achieve this in a cost-effective manner, we propose an approach that uses data derived from radiology reports to enhance the attention of a DL model toward lesion characteristics. The current study aimed to facilitate lesion detection at brain MRI using this approach, while generating interpretable attention maps without the need for manual image segmentation during model training. [Introduction (7), CC-BY 4.0]