This retrospective study used only pseudo-anonymized data consisting of images and clinical data. Ethics approval by the appropriate Institutional Review Boards and written consent to participate was obtained by the MEDIOLA team, sponsored by AstraZeneca, with trial registered with ClinicalTrials.gov, NCT02734004, and all patients provided written informed consent. More details, including the start of participant recruitment, can be found in a previous publication in Drew et al. (17).

Patients in this trial received olaparib (PARPi) monotherapy for the first four weeks, and then a combination of olaparib and durvalumab (ICI) until disease progression or intolerable toxicity. Imaging was performed with the use of CT or Magnetic Resonance Imaging (MRI) at baseline, four weeks (upon completion of PARPi monotherapy) and every eight weeks thereafter. Response outcomes were assessed using RECIST (version 1.1) Eisenhauer et al. (8) by MEDIOLA clinical investigators. Monthly serum CA-125 concentration values were provided to assess disease response and progression through this marker Rustin et al. (18).

Patients included in this imaging-based subset study (N = 20, identified by alphabetical characters A-T for the remaining of this work) presented germline BRCA-mutated, platinum-sensitive relapsed ovarian cancer and their Contrast Enhanced CT (CECT) scans of the thorax, abdomen and pelvis were available at baseline (t₀), week 4 (t₁) and week 12 (t₂, first PARPi-ICI combination assessment). Thirteen (65%) patients carried germline BRCA1 mutations and seven (35%) had BRCA2 mutations. All patients had received previous lines of chemotherapy, with nine (45%) patients having received one previous line, four (20%) patients two previous lines and six patients (30%) three or more previous lines. For one patient, the number of previous lines of chemotherapy was unknown.

For the purpose of the analyses in this study, different response assessments were considered. As a result, every patient was classified as responder or non-responder based on the following four categories:

All lesions identified at baseline (t₀), week 4 (t₁) and week 12 (t₂) CECT scans were manually segmented and cross-validated by two Radiologists in training (VB and DH, both with 6 years’ experience) using the Open Health Imaging Foundation viewer (Open Health Imaging Foundation, Massachusetts Institute of Technology, Cambridge, MA, USA) via its plugin to XNAT, hosted at the local node of the repository established by the CRUK National Cancer Imaging Translational Accelerator (NCITA, https://ncita.org.uk) Doran et al. (25). Lesions were labelled using a numerical code system –within parentheses– upon location:

Note that the most common sites for ovarian cancer patients (namely omentum and pelvic/ovarian) were only found infrequently (pelvis) or not at all (omentum) in this cohort, as patients had received prior chemotherapy and debulking surgery.

Scans in the original Digital Imaging and Communications in Medicine (DICOM, http://medical.nema.org/) format and segmentation files in the DICOM-RT Law and Liu (26) format were converted to the Neuroimaging Informatics Technology Initiative (NIfTI) format Cox et al. (27) using custom software written in MATLAB (The Mathworks Inc., Natick, MA, USA) version R2019b for subsequent feature extraction.

Anatomical networks of the metastases were constructed to represent the distribution of the disease. Each node in the network corresponds to a metastatic site (See Supplementary Figure S1 ) and they are interconnected for every patient and time point. Information from multiple patients was combined into a single network by proportionally scaling the size of the nodes and the thickness of the edges according to the number of patients that were found with those.

The number of edges (patients with disease in the sites on both ends) was taken as a measure of the complexity of the networks. Additionally, anatomical dissemination was also quantified by the total number of lesions (individual fully-connected volumes within each site) and two distance measurements: total intersite distance (i.e. the sum of all Euclidean distances between site centres of mass in the image space) and the maximum distance between two sites Cottereau et al. (28).

Complementarily, co-occurrence matrices were created to illustrate the concomitant presence of sites within patients and were of use to identify pairs of sites with higher likelihood to appear at the same time.

Three first-order radiomic features that quantify heterogeneity (namely energy, entropy and uniformity) for every site, patient and time point were extracted using the open-source Python package PyRadiomics version 3.0.1 Van Griethuysen et al. (29). Higher order radiomic features were not taken into account given the small sample size and the less obvious interpretability of such features.

The pre-processing included re-sampling the images to the average voxel size over the range of values of the cohort (0.801×0.801×3.764) mm using Welch sinc interpolation method to minimise effects of different reconstructed voxel sizes Escudero Sanchez et al. (30).

The volume of each site for every patient and time point was used to evaluate volumetric treatment response. The effect of the different treatment arms to the total volume was compared by extracting the relative change in volumes between t₀ and t₁ (first arm, four weeks of PARPi only) and t₁ and t₂ (second arm, eight weeks of PARPi + ICI). The peritoneal and non-peritoneal sites (see list of sites per region above) contribution to total volumetric burden –i.e. the sum of volumes for all sites– was also calculated.

For every patient, the values of the radiomic features Energy, Entropy and Uniformity were extracted for all sites independently, and the median was used to quantify their overall value. Additionally, as an alternative measurement of the spatial heterogeneity presented by these variables across different sites, their range values were also calculated. Moreover, a summed value of each feature was obtained by adding up its value at every site. When the whole cohort was analysed over time, one patient (patient A) was excluded as no lesions were identified at t₂ and the Wilcoxon signed-rank tests on paired samples cannot be performed under the absence of measures.

Descriptive and inferential statistics were applied to imaging-derived features. Nonparametric tests were conducted and considered statistically significant if p ≤0.05. For paired samples (i.e. same patient, different time point or same patient, different lesion site), Wilcoxon signed-rank tests on paired samples were used. A Kruskal-Wallis test was performed to assess differences between defined groups. All statistical analyses were conducted using R (31) version 4.2.1.

Patients were enrolled in the clinical trial for an average of 64 weeks (range 18–126 weeks, median 49 weeks). Separated by a cut-off at one year, ten patients were short-term enrollees (STE) and ten were long-term enrollees (LTE) ( Figure 1 ). The average Progression Free Survival (PFS) time was 53 weeks (range 4-124, median 44 weeks), with 9 patients below nine months (short responders, SR) and 11 above (long responders, LR). According to their RECIST assessments, five patients were early responders (ER) and 15 non-responders (NR) at four weeks, and 14 achieved partial or complete response over the entire trial versus 6 who did not, following the Best Overall Response (BOR) RECIST assessment criteria.

None of the patients had progressed by CA-125 assessment within the first 12 weeks, hence CA-125 was not used as a response measurement in further analyses.

Figure 2 compares the different response assessment measurements analysed in this study. A comparison with the earliest measurement available in the trial, which is RECIST response at four weeks, is presented in Figure 2A . The confusion matrices clearly show that the early assessment obtained with RECIST is not a good indicator of long-term patient response to treatment. In particular, it was found that many patients who did not respond according to RECIST at 4 weeks (15 patients) did achieve response during the overall treatment (BOR, 9 patients - 60%), were long time enrollees in the trial (LTE, 8 patients - 53%) or had at least 9 months of PFS (LR, 9 patients - 60%).

The relative change in total volume between t₀ and t₁ was also studied as an early assessment of response, motivated by the results observed in Figure 2B , which indicate that several patients deemed as NR at 4 weeks based on RECIST measurements presented a large change in total volume. Establishing a cut-off at 30% in volumetric changes [based on the same relative change of 30% used in 2D by RECIST Eisenhauer et al. (8)] to classify patients as responders or non-responders, a better correlation between this early response measurement and long-term response was found, as observed in Figure 2C .

Lesions were identified and segmented in 15 distinct anatomical sites across the cohort ( Figure 3 ). All patients had metastatic disease at enrolment, with a median of four metastatic sites (range 1-10) per patient at all time-points; being the mesentery the site with the highest occurrence, present in ten (50%) patients at baseline. Non-significant differences were found (Kruskal-Wallis testing, p > 0.05) when comparing, for any of the response assessment classifications, i) the median total number of sites or ii) the median number of peritoneal sites, lymph nodes or other sites at any time point ( Supplementary Table S1 ). As observed in Figure 3B , after the first treatment arm (PARPi only) disease in some peritoneal sites only disappeared: the mesentery disappeared for two patients and, for a third patient, lesions in both LPG and peritoneum (other) disappeared. Additional sites, including several LN disappeared for 7 patients over the second arm treatment (PARPi + ICI), whilst only in one patient a new site (LPG) appeared.

Co-occurrence matrices ( Figure 3C ) were created to identify clusters of sites that had the tendency to appear together. It was observed that over 75% of the patients (85% at t₀ and t₁, and 75% at t₂) that presented lesions in any of the lymph nodes were found to have lesions in other LN too.

As a novel approach to measure spatial heterogeneity and changes thereof, anatomical networks were constructed for each patient and time point ( Figure 4 ). The number of edges of every patient can be found in Supplementary Figure S2 . While the median number of edges at every time point remained constant at 12 edges, the mean decreased (21 at t₀, 19 at t₁ and 14 at t₂). The minimum value of the range was 0 for the three time points (given that four patients in the cohort had one site only), but the maximum decreased from 90 at t₀ and t₁ to 56 at t₂. Figure 5 compares the number of edges for every time point and response measurement assessment. The median number of edges is lower for the responders at all time points for every response assessment measurements; however, the differences are not significant (Kruskal-Wallis testing, p > 0.05). Moreover, when assessing longitudinal changes in the number of edges for each classification, only BOR responders present a significant decrease in the median number of edges at t₂ when compared to baseline (paired samples Wilcoxon signed-rank tests, p < 0.05). Nevertheless, the following trends were observed in the average number of edges per patient ( Figure 4 ): there is typically a decrease in all categories, smaller during the first treatment arm, but more noticeable in the second arm, especially for responders in the long term measurements (LTE and ≥9 months PFS) for which the value halves.

Firstly, anatomical dissemination was quantified by the total number of individual volumes or lesions ( Supplementary Table S2 ). The median number of lesions for the whole cohort at baseline was 14.5 (range 1-58), 11.5 lesions (1-64) at week 4 and 9 lesions (0-49) at week 12. The number of lesions significantly decreased for the whole cohort from t₀ to t₂ (paired samples Wilcoxon signed-rank tests, p = 0.02) ( Supplementary Figure S6 ). Kruskal-Wallis testing was used to evaluate the differences between the different response assessments. For the RECIST assessments, the number of lesions was significantly different only in BOR responders vs non-responders at t₂ (p =0.02). In case of PFS assessments, the number of lesions between SR and LR were significantly different at every time point (p =0.05, p =0.03 and p =0.01, respectively) and for LTE vs STE the differences were significant at t₁ and t₂ (p =0.03 and p =0.01, respectively) ( Figure 6A ). To reduce the effect of the number of sites to the number of lesions (Pearson correlation coefficient r =0.7) ( Supplementary Figure S7 ), the ratio of them was also analysed and resulted to be significantly different for LTE vs STE at t₂ (Kruskal-Wallis testing, p =0.03), for PFS LR vs SR at t₁ and t₂ (p =0.02 and p =0.04) and for BOR responders vs non-responders at every time-point (p =0.03, p =0.04 and p =0.01) ( Figure 6B ).

In addition, the sum of intersite distance and the maximum distance between two sites were studied. For the whole cohort of patients, no significant differences were observed between any time points except for the maximum distance, the median of which was significantly different between t₀ and t₁ (paired samples Wilcoxon signed-rank test, p =0.03) ( Supplementary Figure S3 ). None of the dissemination distance measurements were significantly different when comparing groups ER vs NR, R vs NR by BOR, LTE vs STE or SR vs LR by PFS (Kruskal-Wallis testing, p > 0.05) ( Supplementary Figures S4 , S5 ).

Total volumetric burden decreased significantly for all patients as treatment advanced (paired samples Wilcoxon signed-rank tests, p < 0.05) ( Figure 7A , Supplementary Figures S8 , S9 ); however, the percentage of disease confined in the peritoneum did not significantly differ at any of the time points (paired samples Wilcoxon signed-rank tests, p > 0.05) ( Figure 7B ). Most patients total volumetric burden (13 patients, 65%) decreased both over the first four weeks of PARPi monotherapy (median -29%, range -76%–+48% from t₀) and over the following eight weeks of combined therapy (median -28%, range -112%–+186% from t₁) ( Figure 7C ). Only two patients (P and S) presented a decrease in volume during the first treatment arm followed by an increase in volume in the second arm. For many patients instead (N = 14), the volume decreased over both treatment arms and therefore appear in the third quadrant when comparing the change in volume between the two therapy arms ( Figures 7D, E ). Moreover, although for 13 of these patients the major contribution to the total volume decrease over the 12 weeks happened between t₁ and t₂, two important caveats are to be taken into account: i) that initial volumes were significantly lower at t₁ (see Figure 7A ) and ii) that patients were in combined therapy for double the amount of time. To account for such differences in Figures 7D, E , the changes in volume presented are relative to the initial volume in that range, and the diagonal (around which most of the cases appear), is drawn to compensate for the double amount of time in the second arm.

Volumetric changes over time were also assessed against each traditional treatment response assessment ( Figure 8A ). While the median volume is lower in the responder group for every comparison assessment and time point, the difference is only significant between classes when assessing LTE vs STE (Kruskal-Wallis testing, p < 0.05). Responders by enrollment time, PFS and BOR show a significant decrease over time (paired sample Wilcoxon signed-rank tests, p < 0.05).

The volumetric distribution inside and outside the peritoneal cavity was also studied for all the response assessment classifications and no statistically changes were found neither between classes nor over time ( Figure 8B ). However, for responders according to RECIST assessment at 4 weeks, their disease was either exclusively located outside or inside the peritoneal cavity, with the exception of one patient with a relative volumetric burden in the peritoneal cavity of just 1.05%.

Median, summed and range values of the first-order radiomic features Energy, Entropy and Uniformity for all patients were analysed and presented in Figure 9A . Regarding Energy, changes observed in the second treatment arm and over the twelve weeks of therapy analysed in this study (from t₀ to t₂) were significant, but they were not for the first arm. Changes were less pronounced, and therefore significant, for the other two features. In the case of Entropy, the only significant changes observed correspond to the median value during the second arm and from t₀ to t₂. Similarly, changes were significant for Uniformity only during the second arm for its median value, and from t₀ to t₂ for median and range.

Figure 9B presents the comparison of the median values of Energy, Entropy and Uniformity for responders vs non-responders for the four response assessment measurements considered (see Supplementary Figure S10 for the summed values and Supplementary Figure S11 for the comparison of the range values). Significant differences were found at t₀ and t₁ for LTE vs STE and for PFS responders vs non-responders at t₁ using median Energy. Differences between responders and non-responders are also observed in the distribution of values of Entropy and Uniformity; however, they were not significant.