Company Spotlight: HistoSonics Planning Tool - AI Mapping for Histotripsy

Sam: Welcome to Smart Summaries. Right at the top, we want to remind you that Maya and I are AI-generated voices, and today we are bringing you a summary of a public document.

Maya: Today's document is the FDA 510(k) summary for the HistoSonics Planning Tool, published by the U.S. Food and Drug Administration on May 29, 2026. And the most interesting finding here is that this software can automatically segment the liver with a mean accuracy of 93% to help doctors map out the acoustic pathway for a therapy called histotripsy.

Sam: This is one of our Company Spotlight episodes. We are focusing entirely on one product, but we want to be very clear: this episode is not sponsored. We are independent, and we are basing this conversation purely on the FDA's public 510(k) summary, not on any marketing material from the company.

Maya: Exactly. If it is not in the FDA summary, we are not going to talk about it. So let us start with what this device actually is and the indications for use. The summary describes the HistoSonics Planning Tool as a stand-alone software product that allows physicians to visualize and compare CT and MRI imaging data.

Sam: And it says the display, annotation, and volume rendering aids in histotripsy procedures conducted using HistoSonics Histotripsy Systems. But it is very careful to state that the software is not intended for diagnosis.

Maya: That is a crucial boundary for regulatory clearance. The software is used to evaluate images in the region of interest only for patients already deemed suitable for histotripsy by their treating physician. It operates on a local laptop application or through a secure cloud-based platform.

Sam: So practically speaking, what happens in the software? How does it actually assist the physician?

Maya: The document explains that the tool imports MR or CT studies and allows the user to simulate the patient's position. It overlays simulated treatment planning elements. The physician uses it to localize the target, define the planned treatment volume, and determine the optimal acoustic pathway within the physical constraints of the setup.

Sam: I noticed the summary mentions checking for excessive blockage. What does that mean in this context?

Maya: It means the software allows physicians to check whether there is anything within the acoustic field that would prevent the treatment from reaching the target. It also assesses setup feasibility by identifying potential limitations. For example, it checks if the physical space required for the treatment head, the reservoir, and the patient's position would result in unavoidable collisions.

Sam: That sounds incredibly practical for the operating room. Just mapping out the physical space so equipment does not literally bump into things or the patient. The summary actually claims this has the potential to reduce the overall time the patient is in the procedure room or under anesthetic.

Maya: Yes, by moving these key aspects of the procedure workflow to the planning phase before the patient is even treated. But there is a very important technical note here: the document explicitly states that the HistoSonics Planning Tool and the histotripsy system do not interact or share data.

Sam: Wait, really? They do not share data at all?

Maya: Correct. In the comparison chart under software features, it says export of plans to the patient record is not incorporated. So it is truly a standalone visualization and planning aid.

Sam: Let us talk about how it got cleared. This is a 510(k) summary, which means they had to prove the device is substantially equivalent to a legally marketed predicate device. What was the predicate here?

Maya: The predicate device is the Emprint Visualization System, cleared under 510(k) K192038, manufactured by Covidien llc.

Sam: But the document notes a pretty big difference. The Emprint system is an aid for microwave ablation, whereas the HistoSonics tool is for histotripsy. How can they be substantially equivalent if the therapies are different?

Maya: Because they are both classified under the same regulation for a Medical image management and processing system. The summary argues that they rely on the same principle of operation and fundamental technology, displaying and rendering DICOM data in 3-D. The FDA accepted that the performance characteristics differ only due to the underlying technology of histotripsy versus microwave ablation.

Sam: Right, the FDA Product Code for the predicate is LLZ, but for the subject device it is QTZ, though they share the same regulation number, 21 CFR Part 892.2050. But there is another major technological difference mentioned in the document.

Maya: Yes, the primary technological difference is that the HistoSonics tool allows for automatic segmentation of key organs. It includes algorithms to segment the liver, kidney, lungs, ribs, and skin. All of these models, except for the skin, were developed using machine learning.

Sam: This is where we need to look at the data. For companies building or buying AI, the training and validation datasets are everything. What does the summary tell us about how these models were trained?

Maya: The training dataset was sourced from six publicly available datasets and one proprietary dataset representing over 30 institutions across the United States, Europe, and China. The six public datasets combined had over 3000 scans, with an approximate 6:1 CT to MR ratio.

Sam: That sounds like a robust mix. It mentions imaging was performed across more than 22 scanner models from Philips, Siemens, GE, and Toshiba or Canon. The scan parameters included 1.5T and 3.0T field strengths for MR acquisitions and 120 kVp tube voltage for CT.

Maya: They also included a wide demographic spread. Patient age across the combined datasets ranged from 14 to 100 years, and the gender was approximately a 2:1 male to female ratio when reported. The data included patients with oncological disease, routine cancer screening, healthy subjects, and unknown pathology.

Sam: Okay, that is the training data. But how was it validated? Because the summary is very specific about the evaluation datasets being independent from the training data.

Maya: The validation data came from two public datasets called CRLM and CirrMRI600+, plus a proprietary dataset. And this is where we need to look closely at the numbers. The CRLM dataset includes a total of 197 patients, with a mean age of 61 years. But for the model evaluation, they used a selected subset of 10 patients.

Sam: Wait, they only used 10 patients from that dataset for the evaluation?

Maya: Yes. Those 10 patients were evenly split, 5 male and 5 female, with a mean age of 56 years, ranging from 33 to 85. And here is an important detail: the data was collected from one facility from 1991 to 1995, utilizing Lightspeed 16 and VCT models from GE Healthcare.

Sam: From 1991 to 1995? That is quite old imaging data. And the pathology numbers in that dataset are very specific. It says 58% of the total dataset had multiple metastases, with a maximal tumor size averaging 3.5 cm.

Maya: It also notes the median percentage of necrosis was 30%, ranging from 0 to 90%, and the median percentage of fibrosis was 10%, ranging from 0 to 100%. These pathology details matter because necrosis and fibrosis can change how tissue looks on a scan and potentially how acoustic energy interacts with it, though the document only links this to the imaging segmentation.

Sam: What about the second public dataset, the CirrMRI600+?

Maya: That dataset comprises 339 patients, but again, for model evaluation, 10 patients were selected. Balanced 5 male and 5 female, mean age 53. Imaging was performed at one Turkish site using Philips and Siemens scanners. It included healthy subjects and patients with cirrhosis.

Sam: And the proprietary dataset?

Maya: The proprietary HistoSonics dataset used 55 patients, ages 36 to 81, with a mean of 61. This group had a mix of HCC tumors and liver metastasis. Data was collected at 10 sites in the US and Europe between April 2021 and January 2023. The sex distribution was unknown but noted as 50/50 in the study.

Sam: So you have 10 from the first set, 10 from the second, and 55 from the proprietary set. But the summary mentions a ground truth established through manual segmentation of 78 scans. How did they establish that ground truth?

Maya: The document states that three sets of independent annotations were performed by trained annotators and overseen by radiologists. Those three sets were then merged using the STAPLE algorithm into one consensus ground truth annotation.

Sam: The STAPLE algorithm. The document actually references a 2004 paper by Warfield et al. for that. Once they had that ground truth, how did the AI models actually perform against it?

Maya: They used the Dice coefficient to measure accuracy. The acceptance criteria required a Dice score of 70% or higher in 70% of the cases. Each segmentation model exceeded this. The Liver model reached a mean of 93%, Kidney hit 89%, and Lung achieved 92%.

Sam: What about the ribs? Because ribs are notoriously tricky and definitely block acoustic pathways.

Maya: For the ribs, the evaluation was based on geometric comparison. The acceptance criteria was being within 5mm of the inter-annotator average surface distance. The summary states the Rib segmentation model exceeded this with a Mean 92%.

Sam: And did they break those results down by different patient types to check for bias?

Maya: Yes, subgroup error analysis was conducted across ages 24 to 85, gender at a 50/50 ratio, geographic regions in the US and Europe, and imaging modality with a 2:1 CT to MR ratio. The tool met the criteria with no statistically relevant differences, which demonstrates the robustness of the model.

Sam: That is solid performance for a planning tool. But let us talk about limitations and what companies buying or building this kind of AI should take away from this FDA document. The biggest thing that jumps out is the retrospective nature of the testing.

Maya: Absolutely. Under the clinical study data section, the document plainly states that animal studies and clinical studies in human subjects were not required. The device performance was validated entirely through system level software testing. This is standard for a Class II SaMD, or Software as a Medical Device, but buyers need to know they are looking at retrospective validation, not a prospective clinical trial.

Sam: Another limitation to note is that these machine learning models are static. The document specifically says the device does not incorporate adaptive machine-learning. It will not learn or change on its own once installed at a hospital.

Maya: Right, which aligns with current FDA software life cycle processes. Speaking of which, the performance testing demonstrated compliance with International Standard IEC 62304:2015 for medical device software, and NEMA PS 3.1-3.20:2024e for DICOM. They also provided information demonstrating compliance with section 525B of the FD&C Act for cybersecurity.

Sam: This gives a very clear blueprint for health-tech developers. You can get an AI-powered planning tool cleared by demonstrating substantial equivalence to a non-AI predicate, provided you do rigorous ground-truth validation with radiologists and standard algorithms like STAPLE, even if your validation datasets are relatively small subsets.

Maya: Exactly. From a clinical perspective, if this tool successfully automates the segmentation of the liver and ribs with over 89% accuracy, it could legitimately streamline the setup for histotripsy, reducing the guesswork of whether a tumor is physically accessible before the patient is brought into the room.

Sam: That is a great takeaway. To recap, the FDA cleared the HistoSonics Planning Tool as a Class II device to aid in visualizing and planning histotripsy procedures. Its static machine learning models segment key organs with high accuracy based on retrospective data, and it was found substantially equivalent to an older visualization system used for microwave ablation.

Maya: As always, you can find a link to the full FDA 510(k) summary in our show notes. And remember, we are AI-generated voices, and this podcast is for informational purposes only. It is not medical advice.