Data from 20 patients treated for invasive breast carcinoma between November 2021 and December 2022 were retrospectively selected. Patients were included regardless of age, histology, tumor grade, surgical treatment (lumpectomy or mastectomy), or neoadjuvant chemotherapy. The 20 patients were stratified into four groups: right breast, left breast, right chest wall, and left chest wall. Patient characteristics are described in Table 1 .

All patients were previously treated with TrueBeam in volumetric-modulated arc therapy (VMAT) irradiation. Patients underwent computed tomography (CT) imaging (GE Optima CT580, General Electric Healthcare, Waukesha, WI, USA) with a 2.5-mm slice thickness in a supine position during free breathing and with both arms over the head with personalized foam cushions. In this study, we chose to replan the simulated treatment using an intensity-modulated radiotherapy (IMRT) beam geometry rather than VMAT because of the longer optimization and calculation times with VMAT, which we considered unacceptable for daily adaptive sessions. For this reason, IMRT plans with 13 fields and 6-MV flattening filter-free (FFF) beams were designed for each patient using the Eclipse treatment planning system (v15.6, Varian Medical Systems, Palo Alto, CA, USA) and imported into the Ethos^® solution. Prescriptions for target volumes were 52.2 Gy for the tumor bed (boost) and 42.3 Gy for the breast, the internal mammary chain (IMC), and the clavicular lymph nodes (CLNs) in 18 fractions. CTV–PTV margins were set at 2 mm for all locations except the IMC, for which the margin is 5 mm. The dose constraints for the CTVs and the organs at risk (OARs) are shown in Table 2 .

The dose prescription, PTV margins, and dose constraints were derived from the clinical trial Adaptive Radiotherapy in Hypersensitive Patients and High Locoregional Risk Breast Cancer With ETHOS Technology (SAHARA-04) (18).

The Ethos workflow for breast cancer was reproduced using a Varian Ethos emulator (v1.1, Varian Medical Systems, Palo Alto, CA, USA). First, the cone beam computed tomography (CBCT) scan is imported into the emulator, and the synthetic CT (sCT) is designed using DIR to create the deformation matrix. However, during an adaptive couch session, the sCT was not displayed. Only the CBCT scan was visible. As recommended by Varian (Ethos manual), the body and bone contours (resulting from the synthetic CT) shown on the CBCT were verified by the user for each session to ensure the quality of the synthetic CT. Next, the AI generates the contours of the influencer structures (also known as organs that influence the shape and position of the target), namely, the right and left breasts (or chest walls), both lungs, and the heart. The contours of the target volumes are then generated by elastic or rigid registration, according to the user’s choice, to define the IMC and CLNs. The target propagation is based on structure-guided deformations (resulting from the influencer and bone structures generated in the previous step). The CTV_Breast and CTV_Chestwall are derived structures from the breast and chest wall (which are influencer structures) excluding the 5 mm beneath the skin. The CTV_Boost is a derived structure resulting from an expansion around the surgical fiducials (defined as target volume). During the workflow, the volumes created by Ethos were not edited in order to evaluate the performance of the system. Finally, the adapted and scheduled plans are generated. The scheduled plan is the initial plan recalculated on the sCT, while the adapted plan is the initial plan re-optimized and recalculated on the image of the day.

For each patient, five extended CBCT scans performed initially for their treatment were randomly selected in order to simulate five adaptive sessions.

To evaluate the performance of the Ethos in terms of contour accuracy, the generated structures (influencers and target volumes) were compared with the contours manually drawn on the CBCT scans. Two physicians performed and shared the delineations before the study began. The ESTRO consensus guidelines (19, 20) were used to delineate target volumes, breast/wall, and axillary (Berg I); subclavicular (Berg II, III) and supraclavicular (Berg IV) lymph nodes (Nodes hereafter); and IMC. Organs at risk were delineated following French RecoRad 2022 (21) recommendations using TheraPanacea software (22). Structures derived from influencers and target volumes (CTV_Breast/Chestwall and CTV_Boost) were excluded from this part.

Each CBCT was used twice in the emulator: the first time to recover the target volumes by elastic propagation and the second time by rigid propagation. The generated contours by Ethos were evaluated using two metrics: the Dice similarity coefficient (DICE) index and the Hausdorff 95% distance (HD95). These metrics were chosen because of their common use in the literature (23, 24) and their complementary nature. DICE indices were retrieved using Eclipse TPS (v15.6), and HD95 values using 3D Slicer software.

Finally, the repeatability of the AI-generated influencer contours was investigated by repeating the workflow twice and comparing the influencer generation with 3D Slicer.

In order to compare the dosimetry of the adapted and scheduled plans for each fraction, several dose–volume histogram parameters were reported. To study the target volumes (boost, breast/chest wall, IMC, and CLN), the dose received by 95% of the volume (D95%) and the homogeneity index were used. The homogeneity index is defined by the following formula (25). For the OARs, the mean dose was recorded for the heart, contralateral breast, and both lungs; the volume receiving 17 Gy (V17Gy) for the ipsilateral lung and the mean dose for the contralateral breast are criteria derived from the dose constraints in the protocol for the clinical trial (18) following this feasibility study ( Table 2 ).

The adapted and scheduled plans were also compared with the reference plans, defined as the initial plan performed on the simulation CT.

Contour generation times by AI, elastic/rigid propagation (without edition), calculation, and optimization were recorded. These times were compared between workflows for patients who had been treated without and with mastectomy (breast and chest wall, respectively). Average times were also calculated for all patients.

Ethos includes Mobius3D (version 3.1, Varian Medical Systems), an integrated and independent dose calculation quality assurance tool using the treatment delivery log files. However, our institution’s experience is based on portal dosimetry evaluation. Therefore, quality assurance of the adapted plans was performed using the portal device and the independent software EpiQA (v5.1.3, EPIdos, Bratislava, Slovakia). Each adapted plan was exported to Eclipse TPS, and the planar dose was calculated in a homogeneous cube phantom using the AAA algorithm (Eclipse v15.6, Varian Medical Systems, Palo Alto, CA, USA). Data from the delivered plans were acquired using the portal imager on the Ethos accelerator. The evaluation metric was the global gamma pass rate with the criteria 3%/3mm, 2%/2mm, and 1%/1mm and the field area threshold.

Figure 1 shows the DICE values obtained from the comparison between the AI-generated influencer contours and those delineated by the physician. On average, the DICE is always strictly greater than 0.90 and, in the case of the lungs, greater than 0.95. The same applies to the median. The highest value is reached by the right and left lungs, with DICE of 0.96 ± 0.02 and 0.96 ± 0.01, respectively. The lowest is reached by the right breast (0.92 ± 0.03). The right breast has the largest standard deviation, and the right lung has the smallest. Regarding the comparison between rigid and elastic propagations, the average DICE resulting from contours generated by rigid propagation are larger than those for contours generated by elastic propagation: they are 0.8 ± 0.1 and 0.46 ± 0.17 for rigid for the CLN and IMC, respectively, versus 0.76 ± 0.06 and 0.34 ± 0.14 in elastic, again for the CLN and IMC, respectively.

Figure 2 shows the HD95 values comparing the AI-generated influencer contours with the clinician contours. The mean HD95 does not exceed 10 mm for all influencers. The left breast showed the lowest HD95 and standard deviation with 4.4 ± 1.7 mm. The heart had the highest mean HD95 of 8.8 mm and also the highest variability, with a standard deviation of 4.1 mm. For the right breast and both lungs, HD95 remained between 5 and 6 mm. Similarly, HD95 was lower for rigid contours than for elastic contours in the case of the CLNs: the average was 4.6 ± 1.8 mm in rigid propagation and 5.3 ± 1.6 mm in elastic deformation. For the IMC, the HD95 average was 4.9 ± 2.5 mm for rigid contours and 4.4 ± 2.8 mm for elastic contours. Figure 3 displays an example of a comparison between the physician contour and the rigid propagation contour for the internal mammary chain (left) and the clavicular lymph nodes (right).

The results for the repeatability of influencer generation by AI presented in Table 3 show an average DICE close to 1 and an average HD95 of 0.1 mm.

The workflow can be divided into several stages, each requiring a different amount of time: the time needed to generate contours using AI and propagation and then the time needed to calculate and optimize the treatment plans.

The mean contour generation time per AI was 73 s. Mean target volume contour generation times and associated standard deviations were longer for patients without mastectomy but not significant (85 ± 9 s for breast vs. 76 ± 6 s for chest wall). Calculation times averaged between 50 and 55 s, and optimization times were the longest, averaging between 120 and 130 s ( Table 4 ). The total time for a complete workflow, from creating the influencer to calculating the dose, was 5 to 6 minutes, not including editing.

In view of the results obtained in the previous section, it was decided at the mid-study to evaluate only the plans resulting from rigid propagation, as these were of better quality than those resulting from elastic propagation. The dosimetry for adapted and scheduled plans was evaluated for each fraction, and some dose metrics were evaluated based on the constraints. In addition, some metrics resulting from dose constraints ( Table 2 ) were excluded from the comparison, such as V17Gy or V35Gy for the heart, because the results were too small and not relevant to the study.

Table 5 shows that the homogeneity index remains below 0.1 on average for the adapted plans, except for the breast, where the homogeneity index (HI) is 0.17 ± 0.09. It is greater than or equal to 0.1 on average for scheduled plans, regardless of the target. We also noted that the CTV_Breast has the highest HI and standard deviation among target volumes, whether with adapted or scheduled plans.

In terms of average dose delivered to target volumes, the adapted plans delivered higher D95% than the scheduled plans ( Table 6 ). Furthermore, the D95% for the scheduled plan was inferior to the dose results of the reference plans.

The average dose received by the OARs was similar between the adapted and scheduled plans, in terms of both mean dose received and standard deviations ( Figure 4 ). The largest difference between adapted and scheduled plans was found in the ipsilateral lung location, with a mean dose difference of 0.09 Gy. The same results between adapted and scheduled plans were found for specific organs, as reflected by the median: 8.6 Gy for the ipsilateral lung, 2.5 Gy for the contralateral lung, and 4 Gy for the heart. Furthermore, the dose results obtained for both adapted and scheduled plans were equivalent or inferior to those of the reference plans.

Regardless of the parameter used, the mean gamma index pass rates remained above 95%. For a criterion of 3%/3mm, the gamma index was 99.93% ± 0.38%. The standard deviation was the highest for the 1%/1mm criterion ( Figure 5 ).