Can Advances in Artificial Intelligence Strengthen the Role of Intraoperative Radiotherapy in the Treatment of Cancer?

1. Introduction

Intraoperative irradiation was first performed more than 120 years ago in a woman with endometrial cancer. Since that time, a series of major technological advances have led to the increasing clinical use of intraoperative radiotherapy (IORT) around the world. From the 1930s to the 1950s, low-energy X-rays were used to treat abdominal, thoracic, and head and neck cancers. In the 1960s, Cobalt gamma rays and betatron electrons were used to deliver a single, high dose of radiation, an approach that was first performed at Kyoto University in Japan[1]. In the 1970s, two hospitals in the United States (Howard University Hospital and Massachusetts General Hospital) used conventional linear accelerators for IORT. In the 1990s, dedicated mobile electron linear accelerators and miniaturized low-energy X-ray machines first became commercially available and introduced into routine clinical practice at several centers around the world[1,2]. Later, a new intraoperative modality based on megavoltage electrons was developed. That technique eventually became known as intraoperative electron radiotherapy (IOERT). Although both kilovoltage X-ray IORT and IOERT have been shown to yield excellent treatment outcomes, there is an intense, ongoing debate in the field of radiation oncology among experts and opinion leaders with regards to which of these two techniques is superior[3].

The International Society of Intraoperative Radiation Therapy (ISIORT) was founded in 1998 to foster sharing of clinical experience among treatment centers and to promote the clinical use of this innovative treatment modality. The European section of the ISIORT (ISIORT-Europe), which was created in 2006, established a database registry in that same year[4,5]. The whose purpose was to collect and record technical and clinical data related to IORT at the affiliated centers. Analyses of the data from the ISIORT registry provide a clear picture of the patterns of care in IORT among the participating centers (n=46, mostly in Europe, but also in Asia and Central America), including the IORT technique, the main tumor types, and patient selection criteria. A presentation at the 2021 meeting of the European Society for Radiotherapy and Oncology (ESTRO) reported results from a study of data from approximately 15,000 patients that showed that breast cancer accounted for the vast majority (81%) of IORT treatments, followed by rectal cancer, sarcoma, prostate, and pancreatic cancers[6].

According to some experts, one disadvantage of both kilovoltage X-ray IORT and megavoltage IOERT is that dose distributions must be determined in real-time (i.e. during treatment), which may lead to inaccuracies. In all other modern radiotherapy treatment modalities, by contrast, the dose distribution is precisely determined prior to treatment. The dose distribution is essential for patient positioning and to ensure accurate clinical decision-making. In IORT/IOERT, it is not possible to determine the dose distribution without 3D images obtained immediately after surgery but before the initiation of IORT. Moreover, in the operating room, there is only a short window of time to perform dose optimization following surgery. However, computed tomography (CT) scanners are now available in the operating rooms at a growing number of centers, thus overcoming this important obstacle to dose distribution optimization. In the near future, advances in computational capacity and in deep learning models (a type of artificial intelligence [AI]), are expected to allow medical physicists to quickly and accurately calculate dose distributions, even under the demanding conditions of the operating room.

In this context, the aims of the present article are 1) to describe the most important clinical data for IORT and IOERT) and 2) to explain how advances in computational capacity and large imaging datasets will permit faster and more reliable processing for segmentation (contouring), dose distribution calculations, and evaluation. In turn, these advances will allow for better clinical decision-making based on the unique properties of the specific tumor and individual patient. These datasets could potentially also be reused to help in developing treatment plans for external radiotherapy, if necessary.

2. Main clinical achievements of IORT

The main rationale for IORT is that it only requires a single, high dose of radiation while simultaneously limiting radiation exposure to the surrounding structures, which can be displaced during surgery to avoid unnecessary irradiation. Moreover, the energy beams (especially electron beams) can be easily set up to minimize the dose beyond the target. In addition to this advantage in spatial selectivity, in recent years another highly intriguing biological rationale has emerged: several studies have shown a positive influence on the local microenvironment, which can improve local immunity to potentially reduce the risk of local recurrence[7,8,9,10].

IORT can be used in a wide range of cancers, including breast and rectal cancer, sarcoma, pancreatic cancer, and others (For selected studies see Table 1). Numerous centers now use IORT as a routine treatment for breast cancer. In some centers, IORT is administered as the single radiation modality, which has the advantage of reducing overall local treatment time, thus improving quality of life, while also reducing waiting lists. In other centers, IORT is administered as a boost in combination with EBRT to improve local control (LC) in high-risk patients.

IORT in Breast Cancer

The feasibility and the efficacy of single-shot IORT, a type of partial breast irradiation (PBI), has been studied in well-selected patients[11,12]. The ESTRO and the American Society for Radiation Oncology (ASTRO) have separately developed criteria to guide treatment indications for PBI and accelerated PBI (APBI)[13,14,15]. Numerous studies have described the use of IORT—whether delivered through electrons or low-energy X-rays—in the treatment of breast cancer, either as a single shot modality or combined with EBRT.

The ELIOT trial compared IOERT (21 Gy) to whole breast irradiation (WBI) in women with early breast cancer. Although in-breast tumor recurrence (IBTR) was higher in the IOERT group, there were no significant differences between the groups in overall survival (OS), nor were there any differences in IBTR in the low-risk patients. Based on those findings, the authors concluded that IOERT could be offered to selected patients at low risk of IBTR[16,17]. A few studies have also reported encouraging results with the administration of a single fraction to the surgical bed after local tumor recurrences[18].

Fastner et al. performed a pooled analysis of data from the ISIORT centres, which showed that IOERT administered as a boost for dose intensification reduced local recurrence. In that study, a 10 Gy IOERT boost delivered prior to WBI yielded LC rates that were superior to that obtained in similar studies without IORT[19]. More recently, the ESTRO-ACROP (Advisory Committee for Radiation Oncology Practice) conducted a review of published data on IOERT, concluding that IOERT should be considered as boost technique in higher risk patients. That review also showed that single shot IOERT is a feasible APBI technique for patients with low-risk disease, despite mild acute and late toxicity[18]. The Targit A trial found no differences between IORT (low-energy X-rays) and WBI in terms of local recurrence-free survival, survival without mastectomy, survival without distant relapse, or breast cancer-related mortality[20].

In short, the available clinical data clearly show that IORT—whether delivered with electrons or low-energy X-rays—delivered as PBI, is a safe alternative to EBRT in patients with low-risk breast cancer and as a boost in patients at higher risk of local recurrence.

IORT in Rectal Cancer

In rectal cancer, studies have shown that a single, high dose of IOERT to the tumor bed can improve local control and prevent pelvic relapse in locally-advanced primary tumors. IOERT can also be used to treat recurrent tumors. This is important given that up to 50% of these patients develop metastatic disease, which has a poor prognosis. Several reviews have shown that IOERT improves local control in locally-advanced primary tumors and recurrent tumors[21,22,23,24]. A pooled analysis of 605 patients with locally-advanced rectal cancer (LARC) who underwent multimodal treatment—including IOERT—found low local recurrence rates (12%), even though these patients were considered high risk[25].

To date, only two randomized trials have compared IOERT to other treatment modalities in relatively low risk patients (i.e. those who are unlikely to benefit from IOERT)[26,27]. Dubois et al. conducted a small trial to compare conventional preoperative EBRT to EBRT plus IOERT; however, there were no significant differences in treatment outcomes between the two groups[26]. In 2021, Masaki et al. reported the results of a randomized trial involving patients with LARC. That trial compared oncologic outcomes in patients who underwent surgery with IORT and bilateral pelvic autonomic nerve preservation (treatment arm) versus no IORT and limited nerve preservation (controls). The trial was closed early because the interim analysis showed that distant metastasis-free survival rates were significantly lower in the IORT arm, without significant between-group differences in OS or LC[27].

In the last decade, several systematic reviews and meta-analyses have been performed to clarify the role of IOERT in LARC[28]. In 2020, the ESTRO-ACROP published recommendations for the use of IORT in locally-recurrent rectal cancer[29], based on a comprehensive review of the literature, including a review of 2,843 patients from 21 different studies. Based on the findings of that review, IOERT was recommended in patients with potentially-resectable stage T3-T4 disease after preoperative chemotherapy + EBRT in case of gross residual disease or microscopically positive or close (< 2 – 5 mm) soft tissue margins. The total recommended doses were as follows: R0: 10-12.5 Gy; R1: 12.5-15 Gy; R2: 15-20 Gy. The NCCN guidelines (v. 3.2024) also recommend IORT as a boost in patients with very close or positive margins, particularly in stage T4 or recurrent cancers.

IORT in Sarcomas

IOERT can be considered part of the multimodal treatment of sarcoma due to its ability to deliver a high dose of radiation to the target volume [30]. The potential advantage of IOERT in this setting is that only a limited volume of healthy tissue is exposed to high dose radiation, which is crucial given the negative correlation between irradiation of healthy tissue and late toxicity (e.g. edema, fibrosis and impaired joint function), which can lead to poor functional outcomes[31,32]. An intraoperative boost can be delivered under visual control precisely to the high-risk region, without the need for additional margins to compensate for daily setup uncertainties. Moreover, key structures—including the peripheral nerves, great vessels, and skin—can all be excluded from the target volume, further reducing the risk of late toxicity[33]. Several reviews have shown that IOERT can achieve high control rates in selected cases of soft tissue sarcoma of the trunk and extremities[34,35,36,37,38].

Studies in patients with retroperitoneal sarcomas have consistently shown that preoperative EBRT + IOERT achieves 5-year LC rates ranging from 51% to 89%, which compare favorably to surgery alone and surgery + EBRT [39]. In general, preoperative radiotherapy should be preferred to postoperative radiotherapy due to better LC rates and less toxicity[40]. A recent single-arm, phase I-II dose escalation study was performed in patients with retroperitoneal sarcoma, many of whom had grade (G3) lesions and leiomyosarcoma. Treatment included preoperative IMRT, surgery, and IOERT. Five-year LC and OS rates were 59.6% and 59.5%, respectively[41]. The results of the EORTC STRASS trial showed that preoperative irradiation can improve abdominal recurrence-free survival in patients with primary, well-differentiated liposarcoma and G1-2 dedifferentiated liposarcomas, although only some patients with sarcoma benefitted from this treatment[42].

The main aim of combined radiotherapy (IOERT + EBRT) in patients with sarcoma of the extremities treated with limb-sparing surgery is to achieve local control [43]. In the review by Roeder et al. 5-year LC rates ranged from 82%–97% and treatments that included IOERT consistently achieved excellent limb preservation rates (81%–100%) with good functional outcomes (59%–100%)[39].

IORT in Pancreatic Cancer

IOERT is an innovative therapeutic approach to dose intensification to improve local control in patients with locally-advanced pancreatic cancer (LAPC), one of the most lethal cancer types (5-year OS < 5%) [44]. Numerous reviews have demonstrated that IOERT improves treatment outcomes in pancreatic cancer. However, because IOERT is typically administered as part of a multimodal approach, it is difficult to determine the extent to which this modality influences treatment outcomes[44,45,46].

IOERT can be used in multimodal treatments as a boost strategy for dose intensification in borderline resectable disease or as the only radiation treatment for localized, non-resected, borderline, or post-resection pancreatic cancer[47,48]. IORT has been shown to improve local control in patients undergoing extended resection[49]. IOERT can be used as an individualized, risk-adapted technique to maximize locoregional control through dose escalation[50].

A non-randomized, retrospective study compared IOERT to concurrent chemoradiotherapy (CRT) in patients with LAPC[51]. There were fewer adverse events in the IOERT arm, with a shorter treatment time. Although IOERT was highly feasible, it did not offer any significant advantages over concurrent CRT in terms of survival or tumor control outcomes

The ESTRO IORT Task Force recommendations for IORT in borderline-resected pancreatic cancer and unresected pancreatic cancer were both published in 2020[52,53]. Those articles described the treatment modalities and recommendations for IOERT in patients with close surgical margins or residual disease and in selected cases with unresectable tumors. More recently, Lee et al. evaluated immune response activity in a group of surgically-treated patients with pancreatic cancer who received IORT. Patients treated with IORT showed increased levels of cytokines involved in the PI3K/SMAD pathway in the peritoneal fluid, which are associated with inhibition in the growth, migration, and invasiveness of pancreatic cancer cells[54]. The findings of that study seem to support the hypothesis that IORT activates immune mechanisms that could prevent local recurrence and, potentially, even distant relapse. Yanagi et al. recently compared patients (n=99) diagnosed with non-metastatic pancreatic cancer who underwent one of three different treatment regimens: 1) IOERT + surgery, 2) surgery + postoperative radiotherapy, or 3) radiotherapy alone[55]. LC and OS rates were similar in the two surgical groups, with no complications ≥ G3.

Although current ESMO (European Society for Medical Oncology) guidelines recommend chemotherapy alone for LAPC, data from dose intensification studies suggest that radiotherapy should be considered as an individualized treatment option in selected patients[56]. The results of the CONKO-007 trial showed no significant differences between chemotherapy and CRT in median OS and resection rates; however, the R0 and pathologic complete response rates were both higher in the CRT group[57]. Moreover, OS was longer in the patients who underwent surgical resection. Among long-term survivors, 5-year OS rates were higher in the CRT group (especially the resected subgroup), a finding that suggests that induction chemotherapy followed by radiotherapy could potentially improve outcomes in patients with LAPC.

IORT in other cancers

In addition to the cancer types described above, IORT has also been widely used to treat several other types of cancer, usually administered as a boost in combination with pre- or post-operative radiotherapy. An analysis of the ISIORT registry described the treatment characteristics of IORT in a wide range of cancer types (gastric, esophageal, prostate, gynecological, central nervous system, skull base, head and neck, lung, kidney, and bladder)[5].

3. Technological advances in IOERT planning and delivery

One of the major advantages of IORT is that the radiation is delivered during surgery. Both IORT and IOERT have substantially improved over time through a series of continuous technological advances, including the advent of modern, mobile accelerators for IOERT[58,59,60]. One of the main drawbacks of IORT versus EBRT has long been the lack of high-quality images from advanced imaging techniques (CT, PET-CT, and/or MRI), which are needed to ensure accurate contouring of the volumes of interest (gross tumor volume [GTV], clinical tumor volume [CTV], and organs at risk [OAR]). High-quality images are also needed to accurately calculate the dose distribution. This drawback is especially apparent when compared to the highly advanced computerized treatment planning systems (TPS) that are now in widespread use for EBRT. These modern TPS show the isodose locations inside the body in three dimensions. This allows physicists to adjust the radiation parameters and to instantly visualize the dose distribution, which can be rapidly recalculated to suit the new patient set up[61].

During IORT, radiation oncologists and surgeons work closely together in the operating room to estimate the size and depth of the CTV based of anatomical landmarks in the surgical bed. In some cases, ultrasound probes, which permit visualization of deeper body structures, are used to ensure radiation sparing[62]. However, the accuracy of this process is highly reliant on the clinical experience of the operators, who then select the most appropriate collimator and beam energy (for electron beams). Unfortunately, in IORT, the true dose distribution in the body cannot be visualized at the time of dose application due to the lack of high-quality images, as CT and MRI scanner are not usually available in the operating room. However, the dose distribution could be easily modified, similar to EBRT, if the necessary images were available. In this regard, the introduction of hybrid operating rooms equipped with CT and MRI allows clinicians to obtain real-time images of the patient’s anatomy. While these images are essential to accurately plan IOERT, they are also useful for the surgical intervention. This allows for rapid contouring of the target and non-target structures to create a treatment plan based on real-time data. Clearly, real-time imaging is essential for optimal treatment planning. The integration of navigation systems with IOERT systems could significantly enhance the accuracy and efficiency of collimator setup.

Several reviews have shown that in-room CT imaging increases dose calculation accuracy because the dose distribution is based on a real-time assessment of the patient’s anatomy [63,64] . CT imaging is crucial because some tissues and/or structures are temporarily displaced to avoid the radiation beam. While in-room imaging is a prerequisite to improving IORT, some technical constraints associated with real-time 3D treatment planning (i.e. visualization of the dose distribution for individual patient anatomy) still need to be resolved. To determine the combined effect of IORT and EBRT doses, deformable registration systems are needed to map the dose distribution based on the patient’s anatomical status immediately prior to the second treatment[65,66]. Even after resolving the issues related to differences in terms of tissue arrangement between IORT and EBRT, the distinct biological effects per dose unit in IORT and EBRT—delivered on different time schedules—must be addressed. These considerations should be approached in a manner similar to that used for the combination of brachytherapy and EBRT.

CT scans are required to perform dose calculations in the target location, which must take into account the current clinical and anatomical status of the patient during the operation. An indirect solution is to simulate the applicator position on CT scans taken before the procedure (virtual planning). This approach may be sufficiently accurate, but only if accurate modelling algorithms (a type of AI) can be developed. At present, the most accurate, direct dose calculation method is to perform imaging immediately prior to or after tumor resection[64]. Ideally, the CT scan would be performed in the operating room. However, if in-room scans are not available, the patients could be transferred under anesthesia to another room. The drawback of this approach is that the imaging room must provide a sterile environment (similar to that of the operating room). Clearly, this approach would increase the complexity of the procedure and require more time. A better solution would be to use an in-room imaging system to perform image-guided IOERT, which could include ultrasound, cone-beam CT, or CT[67,68,69].

To accurately represent the patient’s anatomical status, images must be acquired with the collimator in place and positioned as it would be during treatment. This requires a CT scanner with a bore large enough to cover the entire treatment region—including the patient and the applicator in situ—without risking collision with the treatment table or surgical accessories. The CT scanner must also provide images of sufficient quality to allow for real-time electron beam dose calculations. To minimize imaging artefacts, the applicator should be made of non-metallic materials[70].

Target delineation is an essential aspect of IORT to ensure precision. However, contouring can prolong the surgical procedure can could negatively influence the quality of the intervention by allowing fluid to accumulate in the wound, which alters the geometry. In addition, the extra time required could affect the overall performance of the unit by reducing the number of procedures that can be performed in a given time period. This disadvantage could be overcome by using automated contouring methods. Several different methods are available, including advanced algorithms to compare differences in densities and deep learning approaches that use previously compared images as a learning set. The main obstacle to using deep learning (a type of AI) is the need for a reasonably large learning set (i.e. a large number of cases).

Automated segmentation and instant dose distribution in IORT: advances in AI

Computers have long been used to perform dose calculations, although there is always a degree of uncertainty. For this reason, the process might better be described as dose prediction. Even when the calculations are based on well-established formulas, small errors in the input parameters can negatively affect the results. In this regard, an important advantage of modern computers is that they are not limited to pre-set algorithms, but can actually develop their own algorithms or models through “repeated options checks” to predict the development of events. The more data available and the more frequently these patterns are checked against known results (the learning set), the more accurate the results become. The accuracy is highly dependent not only on the size of the learning set and the calculation capacity of the computer, but also on relevance of the learning set used to train the model.

Due to the major technical advancements that have been made in recent years, artificial neural networks (ANN) can now match—and even surpass—human neural networks in terms of accuracy. While ANNs have long had the ability to process previously structured data (selected parameters), they only recent gained the capacity to process raw data, especially images, which represented a gigantic advance in their processing capacity. Deep learning from unstructured data (raw images) is a major development, as it permits images obtained from diverse sources to be used for training (“big data” approach). This process is a clear example of artificial “intelligence”, as it is similar to how humans draw conclusions.

The increased potential to closely integrate computers into decision-making processes has raised concerns about the appropriateness of learning datasets. For instance, if the images used to create the learning dataset do not adequately cover all scenarios, the resulting machine-based segmentation may not be correct. If the human responsible for supervising the proposed segmentation (the physicist or physician) fails to detect this error, the patient may be harmed. Currently, AI can perform some of the steps in the radiotherapy process, including segmentation (OARs and target), deformable image registration, and treatment planning. AI can also be used in image-guided and adaptive radiotherapy and to predict treatment response and determine prognosis[71]. However, all of these applications require large, relevant learning datasets.

There are many benefits to automated segmentation. First, manual segmentation is a highly time-consuming process and automated segmentation can reduce this from hours to mere minutes. Moreover, segmentation is highly dependent on the individual radiation oncologist (high interobserver variability) and can be affected by the number of sections, the complexity of the involved structures, and the availability of support software. One study found that AI-assisted EBRT reduces segmentation time by 65% (5.4 min) and decreases inter-observer variability for the GTV by 32%[72]. These findings are relevant in the surgical setting in which IORT is performed due to the importance of limiting the time required for target delineation following image acquisition. In EBRT, delineation is performed prior to treatment and can be done manually because time constraints are less important than in IORT. Even so, AI-assisted software is increasingly employed in EBRT. In IOERT, contouring must be performed during surgery, with an open wound susceptible to infection. In this setting, time is a crucial factor.

In IORT, the target volume is the surgical bed, with the surrounding healthy tissues temporarily displaced to permit direct irradiation of the target. As a result, the tissue anatomy depends on the time point (i.e. preoperative, intra-operative, and after suturing)[73]. Thus, there is a clear need for automated support to reduce the time needed for contouring, ideally to only a few minutes.

In EBRT, automated segmentation is based on large databases of image atlases. Similar imaging databases are needed for IORT, especially because the shape and volume of the body structures can be substantially altered during surgery. This loss of body integrity alters the relationships between organs and tissues and must be accounted for during IORT.

Deep learning methods, based on raw images acquired pre- and post-resection, may enable AI to be used to automatically segment key structures and to rapidly perform the dose calculation. In this approach, images obtained preoperatively are combined with images acquired immediately after surgery (assuming the operating room has a CT scanner). Over time, we may eventually have a sufficiently large learning set. At present, however, it is not clear exactly how these self-developed patterns—in which the learning set is based on numerous preoperative cases but many fewer postoperative cases—will work. Moreover, the accuracy of these patterns is also uncertain. In most cases, in-room CT and MRI-based imaging is not available, and thus the radiation parameters for IORT are selected without the assistance of these images or isodoses. Under these conditions, AI-based automated segmentation seems to be justified. Moreover, since automated segmentation can be completed in only a few seconds and the dose distribution calculated nearly instantly, the applicator can be placed in the optimal location to avoid irradiating healthy tissues.

The validity of learning datasets in large population settings is an unresolved question because a dataset acquired from one population may not be relevant to a different population due to differences in body build, including height, weight, and other anatomical differences (body mass index, obesity, etc.). In this sense, the dataset should take heterogeneity in these parameters into account, many of which are related to differences in age, sex, and body structure. In the future, it may be possible to use AI to create, prior to treatment, virtual reality-based representations of the patient’s probable anatomical status during IORT. This representation could include expected alterations to the organs and any displacements related to open surgery or minimally-invasive laparoscopic or robotic procedures.

AI can have a large impact on dose distributions in IOERT. Knowledge-based treatment planning, a machine learning model, was first introduced in 2014 and is now widely used in IMRT[74]. In this model, data from prior treatment plans are registered and then used to create a semiautomatic plan, thus significantly reducing the time needed for treatment planning. More recent developments include deep learning methods capable of generating new dose distributions based on previous treatments plans, offering the potential to rapidly deliver and optimize treatment planning, in some cases in a question of minutes or even seconds [75,76][. “Intraplanning”, a technique in which the treatment plan is adapted to the patient’s real-time anatomy[77], may become the standard approach to planning in the next generation of IOERT TPS[77].

AI can also support the management of internal body movements that impact target positioning during radiotherapy. For example, in breast cancer treatments, AI-driven systems can account for the respiratory phase in breath-hold settings or help reduce respiratory motion, thereby minimizing or preventing target displacement during irradiation. Similar strategies can be applied to thoracic and upper abdominal cancers. Irradiation of the lower abdomen typically involves relatively high dose rates, which means treatment is short and bowel movement-related anatomical changes are limited. Similarly, bladder filling is typically managed by catheterization. For cancers located in the head and neck region or the extremities, no significant changes due to internal body movements are expected during IOERT.

AI is currently used in EBRT to help determine prognosis through radiomic analysis of imaging studies. AI could be used in the same way for IORT. It could also be used to help identify the patients most likely to benefit from IORT[71]. Individual radiation sensitivity is also relevant to IORT. Several different molecular methods are used to evaluate resected tissues (including cells from the tumor microenvironment). It is now possible to analyze molecular data from both cancerous and non-cancerous cells, enabling the use of these datasets to train machine learning and deep learning algorithms. Furthermore, these molecular data can be integrated with information on dose distribution and side effects to help personalize IORT. Although researchers are actively developing such predictive models, no validated models are currently available[78,79].

Machine support and automation have long been used in industry—particularly in aviation—and now play a well-established role in radiotherapy, where they can greatly reduce the time needed for clinical decision-making. These tools help medical physicists and physicians to optimize treatment-related decisions and to adapt the treatment to suit the individual patient’s needs. However, as automation takes on a greater role in the decision-making process, there is a growing risk of liability. At present, clinical decisions are evaluated according to the available information and on adherence to well-established good practices. If a clinical decision based on a machine-generated pattern turns out to be incorrect, there may be no protocol to address this type of error. In current clinical practice, machine-generated contouring must be reviewed and approved by a human, thus ensuring adherence to clinical protocols and good clinical practice. However, in the IORT setting, this practice is more challenging due to the real-time nature of this technique, in which there is only a short window of time to assess AI-generated recommendations.

5. Conclusions

Intraoperative radiotherapy—particularly IOERT—plays an important role in radiotherapy, offering numerous well-documented advantages in several types of cancers. These intraoperative techniques are supported by an extensive body of clinical evidence, especially in specific settings such as low-risk breast cancer, where it significantly shortens treatment time. IORT can also be used to deliver a boost in selected cases of rectal and pancreatic cancer, and soft-tissue sarcoma, among other tumor types. Given the current capabilities of modern computer systems, together with recent advances in machine-generated decision-making patterns, it is clear that AI could greatly improve various aspects of IORT (as is currently done in EBRT), most notably real-time segmentation and isodose calculations.

In the near future, AI may be used to help predict radiation-induced complications in IORT by taking into account individual risk factors, such as genetics, environmental factors, and lifestyle habits.