Artificial intelligence in Ultrasound: Pearls and pitfalls in 2024

Keywords

AI - radiomics - ultrasound - ethics

During the last 5 years, artificial intelligence (AI) emerged as a revolutionary tool with significant implications across the healthcare system, heavily influencing scientific research and different medical fields including the disciplines of pathology, oncology and radiology at most.

Ultrasound, a cornerstone of medical diagnostics, is also witnessing a transformative process. Recent advancements in AI tools are in fact starting to radically change this field.

Interestingly, in a context where the reliance on the expertise of sonographers has been often considered as a major limitation, ultrasound represents an ideal candidate to receive significant benefits from the integration of AI, with promises of enhanced diagnostic accuracy, improved workflow efficiency, and expanded access to high-quality care.

Additionally, it should be considered that ultrasound is known to be a useful first diagnostic approach or a screening tool for many medical conditions (e. g. screening for breast lesions, screening for abdominal aortic aneurysm, screening for thyroid nodules, surveillance of patients at risk of hepatocellular carcinoma, screening for carotid or lower limb arteries, atherosclerotic disease, etc.). However, its use cannot be planned widely enough to satisfy such needs that involve very large populations, mainly due to the limited availability and costs of expert medical manpower. Automated use of ultrasound, guided by AI recognition, may have the potential to speed up at least some of such processes reducing the needed manpower.

Altogether, the impact of AI in ultrasound is multifaceted, affecting different phases of the examination, ranging from image acquisition, recognition of abnormalities and interpretation to decision-making and patient outcomes; however, some peculiarities of AI tools may slow down its application in clinical practice and must be known to avoid troublesome situations.

Hereinafter, we will discuss the major benefits correlated to the introduction of different AI tools in ultrasonology and some of the obstacles that need to be overcome.

Enhancing Diagnostic Accuracy

One of the most significant and easier to perceive contributions of AI to ultrasound imaging lies in its ability to enhance diagnostic accuracy. AI algorithms, particularly those based on deep learning and machine learning, have demonstrated remarkable proficiency in image analysis.

In a large retrospective analysis on more than 100 000 people, Gao et al. demonstrated that AI-assisted ultrasound imaging using deep learning techniques enabled to discriminate between malignant and benign ovarian masses with excellent accuracy performances confirmed by an AUC of 0.87 on an external validation cohort [1].

A recent systematic review on the topic confirmed optimal accuracy across different studies with AUC ranging from 0.73 to 0.99; however, it is important to highlight that only 2 out of 37 studies that evaluated the accuracy of the model used an external validation set [2], which is a necessary step before any tool can be brought to clinical practice.

Another fascinating application of AI in ultrasound was explored by Fu et al., the authors conducted a meta-analysis involving 11 005 people which proved that the diagnostic accuracy of ultrasound-based radiomics had sensitivity and specificity of 0.76 and 0.78 for predicting HER2, 0.80, and 0.76 for Ki67, 2 key biomarkers in the clinical management of breast cancer [3].

Again in the field of breast cancer, the addition of an AI analysis improved the diagnostic accuracy of ultrasound, reducing the rate of false-positive cases for cancer [4].

In the oncology field the number of potential applications of AI is impressive [5]. A recent preliminary work claimed ability to improve the capacity of ultrasound to classify metastatic cervical lymphadenopathies into primary cancer sites [6].

As illustrated through examples, AI in general could provide suggestions for disease characterization, especially combining various multiparametric ultrasound approaches. Noteworthy, this might be particularly useful in the instance of patients hosting unknown rare conditions, which might have never been met before by the operator and might not even come to his mind, but they are likely to become recognized once suggested by an AI tool.

Improving Workflow Efficiency

Beyond diagnostic accuracy, AI is also revolutionizing the operational aspects of ultrasound imaging. Traditional ultrasound examinations can be time-consuming and require significant expertise. AI streamlines these processes through automation and intelligent assistance, thereby improving workflow efficiency.

For instance, a regression convolutional neural network designed by Fiorentino et al. [7] to measure the fetal head circumference, a crucial parameter to evaluate fetal growth and development, proved to be able to reduce examination time, inter-clinician variability, and increase diagnostic accuracy [7].

Another key area where AI has made substantial inroads is in the acquisition of ultrasound images. Automated image acquisition systems, guided by AI, can assist clinicians and sonographers in capturing high-quality images consistently. These systems use real-time feedback to ensure optimal probe positioning and image quality, reducing the dependency on the operatorʼs skill level.

This particular aspect was proved in a clever prospective: In a multicenter diagnostic study 8 nurses, with no previous experience of ultrasound imaging, performed an echocardiography on 240 patients using a deep-learning AI–based software based on 5 million examples of the outcome of ultrasonographic probe movement on image quality.

The images were later assessed by 5 expert echocardiographers who judged them of diagnostic quality in 98.8 % of patients for left ventricular size and function, and 92.5 % for right ventricular size [8].

Expanding Access to High-Quality Care

AI’s transformative impact on ultrasound imaging extends to addressing disparities in healthcare access. In many parts of the world, especially in low-resource settings, access to quality ultrasound services is limited due to a shortage of trained personnel and equipment. AI has the potential to bridge this gap by democratizing access to high-quality imaging.

Notably, in a retrospective analysis published in this issue of Ultraschall in der Medizin/European Journal of Ultrasound, Wei et al. used an automatic recognition system using neural network to diagnose the presence of carotid plaques in real-time and synchronously to the acquisition of the image, achieving a remarkable accuracy of 98.5 % [9]. Additionally, using live transmission of ultrasound images, they tested the use of the algorithm to diagnose carotid plaques in ultrasound examinations during their performance as far as more than 1000 km away, showing how distances can be shortened thanks to these tools [9]. Such technologies would allow the potential of delivery of care in the absence of local expertise and also in case of shortage of qualified manpower to assist the examination live from remote.

In this context portable ultrasound devices are increasingly being deployed in remote and underserved areas [10]. These devices are compact, affordable, and easy to use, making them ideal for field conditions. AI assistance in real-time image interpretation could provide instant diagnostic support to healthcare workers who may lack specialized training, enabling timely diagnosis and treatment. AI assistance also allows the recognition of anatomic structures during point-of-care interventional procedures.

This not only enhances diagnostic accuracy, but also fosters local expertise by providing educational feedback to healthcare providers in remote areas.

Challenges and Ethical Considerations

While the progress of AI in ultrasound imaging is undeniably impressive and fast rising, it is not free of challenges and ethical considerations worth to be taken into account.

First of all, ensuring the reliability and generalizability of AI algorithms is of paramount importance. Unfortunately, many AI models are trained on specific datasets, and their performance can vary when applied to different populations or clinical settings. Therefore, external validation and updating of these models are essential to maintain their accuracy and relevance.

Secondly, another key issue is represented by the so-called black box problem. This reflects the inability to understand how deep learning models reach their conclusion, causing potential catastrophic consequences. In order to safely introduce AI tools in clinical practice it is mandatory to generate explanation of black box mechanisms of action, which some authors recently compared to creating explanations in natural sciences or physics [11].

Strictly connected to the lack of understanding of the AI tool there is the accountability issue. Accountability is often defined somehow imprecisely, but it can be simplified as an obligation to inform about, and justify one’s conduct to an authority [12]. This implies both a legal liability, which has to be strictly regulated before the introduction in clinical practice, and moral responsibility to the patients who are going to be influenced by decisions taken thanks to the help of AI tools.

To this end, in order to be certified by regulatory agencies and marketed, specific versions of AI tools should to be finally fixed; otherwise, if a continuous evolution and improvement of a software takes place (theoretically to be regarded as a great opportunity), with no obvious overt communication of the pass from a version to a new one, it could happen that a diagnosis established today could turn into a slightly different diagnosis tomorrow, despite the same images are analyzed, if an automatic update version of the system has occurred, puzzling both operators and patients. This could even raise legal questions for the responsibility in the diagnostic process.

Responsibility of the final diagnosis and possible or not access of patients to information of the AI software utilized to support the final diagnosis are still matter of debate and concern.

Lastly, data privacy remains probably the most significant ethical concern. Data leakage and successful prompt injection attacks to steal data have been reported [13] [14], showing how AI tools are still lag behind in this aspect which is crucial to allow them in clinical practice. Moreover, who “owns” the images that have been elaborated to produce new software which could be needed for new updates is also another serious concern.

Conclusion

The integration of AI into ultrasound imaging represents a paradigm shift in medical diagnostics and patient care. By enhancing diagnostic accuracy, improving workflow efficiency, expanding access to high-quality care, AI is set to redefine the standards of ultrasound imaging. However, it is imperative to address the challenges and ethical considerations associated with AI to ensure its responsible and effective implementation.

As we continue to explore the potential of AI in ultrasound imaging, the collaboration between technologists, clinicians, and policymakers will be crucial. Together, we can harness the power of AI to transform healthcare delivery, ultimately improving patient outcomes and advancing the frontiers of medical science.

Dr. Bernardo Stefanini

Dr. Alice Giamperoli

Dr. Eleonora Terzi

Prof. Fabio Piscaglia

• References

• 1 Gao Y, Zeng S, Xu X. et al. Deep learning-enabled pelvic ultrasound images for accurate diagnosis of ovarian cancer in China: a retrospective, multicentre, diagnostic study. Lancet Digit Health 2022; 4 (03) e179-e187
• 2 Moro F, Ciancia M, Zace D. et al. Role of artificial intelligence applied to ultrasound in gynecology oncology: A systematic review. Int J Cancer
• 3 Fu Y, Zhou J, Li J. Diagnostic performance of ultrasound-based artificial intelligence for predicting key molecular markers in breast cancer: A systematic review and meta-analysis. PLOS ONE 2024; 19 (05) e0303669
• 4 Eun NL, Lee E, Park AY. et al. Artificial intelligence for ultrasound microflow imaging in breast cancer diagnosis. Ultraschall in Med Stuttg Ger 1980 2024; 45 (04) 412-417
• 5 Vetter M, Waldner MJ, Zundler S. et al. Artificial intelligence for the classification of focal liver lesions in ultrasound – a systematic review. Ultraschall in Med. 2023; 44: 395-407
• 6 Zhu Y, Meng Z, Wu H. et al. Deep learning radiomics of multimodal ultrasound for classifying metastatic cervical lymphadenopathy into primary cancer sites: a feasibility study. Ultraschall in Med Stuttg Ger 1980 2024; 45 (03) 305-315
• 7 Fiorentino MC, Moccia S, Capparuccini M. et al. A regression framework to head-circumference delineation from US fetal images. Comput Methods Programs Biomed 2021; 198: 105771
• 8 Narang A, Bae R, Hong H. et al. Utility of a Deep-Learning Algorithm to Guide Novices to Acquire Echocardiograms for Limited Diagnostic Use. JAMA Cardiol 2021; 6 (06) 1-9
• 9 Wei Y, Yang B, Wei L. et al. Real-time carotid plaque recognition from dynamic ultrasound videos based on artificial neural network. Ultraschall in Med Stuttg Ger 1980 2023;
• 10 Piscaglia F, Stefanini B, Calliada F. et al. Ultrasound in clinical enviroments: Where are we standing?. Ultraschall in Med Stuttg Ger 1980 2023; 44 (04) 353-358
• 11 Marcus E, Teuwen J. Artificial intelligence and explanation: How, why, and when to explain black boxes. Eur J Radiol 2024; 173
• 12 Novelli C, Taddeo M, Floridi L. Accountability in artificial intelligence: what it is and how it works. AI Soc 2023;
• 13 Clusmann J, Ferber D, Wiest IC. et al. Prompt Injection Attacks on Large Language Models in Oncology. 2024;
• 14 Samoilenko R. New prompt injection attack on ChatGPT web version. Reckless copy-pasting may lead to serious privacy issues in your chat.

Publication History

Article published online:
06 September 2024