Innovation to Impact - AI and Focused Ultrasound: Part 1

Last December, the Foundation examined the use of artificial intelligence (AI) and machine learning (ML) techniques in focused ultrasound. We’ve refreshed that analysis to understand how these techniques continue to advance the field and drive meaningful real-world results. In this three-part blog series, to be released over the next month, we’ll explore the following: 1) the current state of AI/ML in focused ultrasound; 2) the ecosystem steps that are transforming other therapies and that may likewise accelerate the development and adoption of clinical focused ultrasound over the next few years; and 3) the rapidly developing field of generative AI (gen AI), and the potential impacts of widespread gen AI adoption on focused ultrasound over the coming decade.

Focused Ultrasound + AI/ML Today: The Intersection of Two Rapidly Evolving Fields

The integration of AI/ML techniques into the field of focused ultrasound is advancing rapidly, with the universal goal of enhancing treatment efficacy in the clinic. A review of numerous studies combining AI/ML with focused ultrasound reveals that these techniques are particularly valuable in several areas, including ultrasound image denoising, patient selection and treatment planning, treatment monitoring and image analysis, and treatment assessment and outcome prediction. This blog provides overviews of each area and related links for those who would like to learn more.

Ultrasound Image Denoising
High-intensity focused ultrasound irradiation generates acoustic interference during focused ultrasound treatments, which contaminates the B-mode images used during treatment monitoring. Convolutional neural networks (CNNs), an AI/ML network architecture, have been used to suppress focused ultrasound interference in ultrasound monitoring images. Results from a recent study showed that the use of one-dimensional CNNs achieved an improvement of more than 30 percent in structural similarity to uncontaminated B-mode images. In other words, the image quality was significantly improved, allowing for better medical analysis. These findings suggest that CNNs can serve as a promising tool to denoise ultrasound images, enabling physicians to improve accuracy in analyzing ultrasound images during focused ultrasound treatments.

Patient Selection and Treatment Planning
AI/ML models have been valuable in assisting physicians with patient selection and appropriate treatment planning. In the past, it has been difficult for physicians to determine which patients are best suited for focused ultrasound treatment due to variables, including target vascularity and cellularity. Radiomics is a method that uses AI/ML algorithms to extract key features from medical images, including information about target texture, size, and shape. Information gleaned from radiomics analyses reveals minute differences in target textures that cannot be detected by the human eye.

In patients with uterine fibroids, for example, prediction of the non-perfused volume ratio (NPVR) is important for identifying those who would most benefit from focused ultrasound treatment and for minimizing the risk of treatment failure. A recent study demonstrated that a model combining radiomics features and clinical parameters holds significant potential as a tool for NPVR prediction. This model captured meaningful clinical indicators that predict NPVR, including fat thickness, fibroid volume, fibroid diameter, and distance between fibroids and the skin surface. An approach like this has the potential to transform clinical treatments, enabling physicians to reduce treatment costs and better advise patients on their treatment paths by better understanding those who are most likely to benefit.

In addition to aiding in patient selection, AI/ML methods have also shown promise in enhancing spatial targeting for focused ultrasound treatments. Transcranial focused ultrasound therapy, for example, offers precise thermal ablation for anatomic targets when treating Parkinson’s disease and essential tremor. However, manual preparation for this procedure involves precisely delineating nerve fibers and identifying risk areas in the brain – a process known as fiber tracking – along with establishing safety margins to ensure accurate treatment. Given that the manual process can be quite time-consuming, a team in Germany recently tested whether automated fiber tracking and automated determination of standard treatment coordinates could be done with deep learning and segmentation techniques. Their findings revealed tremendous potential for automation in treatment planning: while manual planning took approximately four hours per patient, automated fiber tracking reduced this time to just six minutes per patient. However, this significant time savings depends on the accuracy of the automated segmentation and still requires expert oversight to ensure precision, particularly in more complex cases.

Treatment Monitoring and Image Analysis
Monitoring during focused ultrasound treatments is crucial to ensure treatment accuracy, safety, and efficacy. While magnetic resonance imaging and ultrasound imaging are methods that are currently used for real-time monitoring during focused ultrasound treatment, both treatments fall short of capturing permanent changes in tissue post-treatment. Researchers recently explored the use of multi-wave photoacoustic (MWPA) imaging for this purpose, as it is better suited to detect the lasting effects of focused ultrasound treatment. In this study, CNNs were used to segment images after MWPA imaging for improved analysis of lesion development in focused ultrasound–treated areas on bovine tissue samples. Results showed that CNN implementation surpassed traditional AI/ML methods in accurately outlining treated areas. This study demonstrates the high performance of deep-learning–based segmentation in monitoring focused ultrasound lesion formation. Furthermore, it highlights how AI/ML methods can lead to improved patient outcomes by providing more accurate and reliable image analysis.

AI-based ultrasound frameworks also hold promise for substantially enhancing image analysis and thus treatment efficacy. In procedures where the objective is to ablate an entire tumor, tumor remnants may remain within small unablated areas after focused ultrasound sonications. With the absence of a precise monitoring system, any tumor residue that is left between ablated areas can increase the risk of tumor resurgence, ultimately necessitating continued treatment. To decrease the risk of tumor resurgence, physicians may also target a larger treatment volume to account for cancer cell residue. However, this method results in a longer treatment time and increases the risk of healthy tissue damage. Alternatively, AI-assisted ultrasound-guided focused ultrasound has shown promise in identifying ablated tissue with high precision in a real-time manner. Segmentation maps have revealed this framework’s ability to accurately determine the volume of ablated regions after focused ultrasound sonication, greatly aiding treatment control and potentially reducing the risk of tumor metastasis.

Treatment Assessment and Outcome Prediction
In addition to providing information before and during focused ultrasound treatments, AI/ML methods can also help provide critical prognostic information after treatments. A recent study explored the potential of deep-learning–based assessments of sarcopenia (muscle wasting) and myosteatosis (fatty muscle) to predict prognoses for patients with advanced pancreatic cancer following focused ultrasound treatment. Researchers used deep learning algorithms to automatically quantify sarcopenia and myosteatosis through CT scan analysis. Both conditions were revealed to be significant indicators of patient outcomes following focused ultrasound treatment; furthermore, higher levels of these measurements pointed toward poorer treatment responses. This study underscores the value of deep learning for advanced analysis of vital treatment parameters, influencing how physicians gather data to personalize treatment plans for patients with particularly poor outcomes.

Conclusions
Incorporating AI/ML technologies into focused ultrasound treatments can improve patient care across many indications. The applicability of AI/ML techniques continues to be explored in numerous studies to better understand how such technologies can address challenges within focused ultrasound treatments. The present landscape reveals that AI/ML methods have the potential to increase the safety, efficacy, and efficiency of focused ultrasound procedures through faster and more precise image analysis. This advancement will benefit physicians and patients alike by enhancing treatment precision and accuracy, thereby reducing the risk of adverse effects and minimizing the need for additional treatments. Together, those of us working in focused ultrasound need to understand and be inspired by the work happening in this space. We further need to adopt these methods where applicable and improve on them where we can. Ultimately, we are collectively responsible for shaping the future of focused ultrasound, and AI/ML provides a significant lever in improving outcomes for patients.

We look forward to learning how you, the focused ultrasound community, are adopting and transforming AI/ML to drive meaningful change for patients. As always, please do not hesitate to share your thoughts.

This blog was written by Rick Hamilton, the Foundation’s chief technology officer and a managing director, and Isha Bhatia, science associate at the Foundation.

References

Yang K, Li Q, Liu H, Zeng Q, Cai D, Xu J, Zhou Y, Tsui PH, Zhou X. Suppressing HIFU interference in ultrasound images using 1D U-Net-based neural networks. Phys Med Biol. 2024 Mar 14;69(7). doi: 10.1088/1361-6560/ad2b95. PMID: 38382109.

Zhou Y, Zhang J, Li C, Chen J, Lv F, Deng Y, Chen S, Du Y, Li F. Prediction of non-perfusion volume ratio for uterine fibroids treated with ultrasound-guided high-intensity focused ultrasound based on MRI radiomics combined with clinical parameters. Biomed Eng Online. 2023 Dec 13;22(1):123. doi: 10.1186/s12938-023-01182-z. PMID: 38093245; PMCID: PMC10717163.

Klein J, Gerken A, Agethen N, Rothlübbers S, Upadhyay N, Purrer V, Schmeel C, Borger V, Kovalevsky M, Rachmilevitch I, Shapira Y, Wüllner U, Jenne J. Automatic planning of MR-guided transcranial focused ultrasound treatment for essential tremor. Front Neuroimaging. 2023 Oct 26;2:1272061. doi: 10.3389/fnimg.2023.1272061. PMID: 37953746; PMCID: PMC10637361.

Wu X, Sanders JL, Dundar MM, Oralkan Ö. Deep-Learning-Based High-Intensity Focused Ultrasound Lesion Segmentation in Multi-Wavelength Photoacoustic Imaging. Bioengineering (Basel). 2023 Sep 8;10(9):1060. doi: 10.3390/bioengineering10091060. PMID: 37760164; PMCID: PMC10526078.

Sadeghi-Goughari M, Rajabzadeh H, Han JW, Kwon HJ. Artificial intelligence-assisted ultrasound-guided focused ultrasound therapy: a feasibility study. Int J Hyperthermia. 2023;40(1):2260127. doi: 10.1080/02656736.2023.2260127. Epub 2023 Sep 25. PMID: 37748776.

Nowak S, Kloth C, Theis M, Marinova M, Attenberger UI, Sprinkart AM, Luetkens JA. Deep learning-based assessment of CT markers of sarcopenia and myosteatosis for outcome assessment in patients with advanced pancreatic cancer after high-intensity focused ultrasound treatment. Eur Radiol. 2024 Jan;34(1):279-286. doi: 10.1007/s00330-023-09974-6. Epub 2023 Aug 12. PMID: 37572195; PMCID: PMC10791981.