VOI analysis
Ten patients from HALO-109-101 and 13 patients from HALO-109-102 have DW-MRI available from multiple scan dates; however, in one patient (HALO-109-102, 102-002-103) ADC could not be computed as b = 0 s/mm2 images were not acquired. Thus, ADC of a total of 40 tumors in 22 patients were analyzed, out of which ADC replicates at baseline were available for 18 patients with a total of 28 tumors. At baseline, ADC from patient 101-003-103 was not included as the b values used for DW-MRI in this subject (0 & 1000 s/mm2) were a deviation from the prescribed study protocol values (0 & 450 s/mm2). Also, ADC at baseline for patient 102-002-113 could not be calculated as the b = 0 s/mm2 images were not acquired. Therefore, changes in ADC at baseline from 20 patients with a total of 31 tumors were analyzed. Further, 19 patients with multiple scan dates of T1w-MRI and DCE-MRI were available (13 from HALO-109-102 and 6 from HALO-109-201); however, in one patient from HALO-109-102 (patient 102-002-102) T1 maps could not be obtained as T1w-MRI were acquired using one FA only. Additionally two tumors in patient 201-007-405 were not clearly visible on T1w-MRI. Therefore, T1 from 18 patients with a total of 26 tumors, and DCE parameters (iAUC, ktrans, vp, and ve) from 19 patients with a total of 29 tumors, were analyzed. Patient 102-003-104 did not have T1w-MRI images at baseline. Patients 102-003-104 and 102-003-117 were left out of the DCE-MRI analysis at baseline due to missing T1w-MRI and DCE images, respectively. Therefore, changes in T1 in 24 tumors from 17 patients and changes in DCE-MRI parameters in 26 tumors from 17 patients were analyzed. In total, 29 patients were analyzed, and a summary of patient data is described in Table 1. Drug pharmacokinetics and survival information per patient are described in Table 1 Supplementary Material A.
Figure 1 shows parameter maps per visit of patient 102-003-105 (tumor 2) with a colorectal cancer metastasis in the liver. At Day 1 following PEGPH20, a relatively uniform decrease in ADC and T1, and an increase in iAUC, ktrans, and ve, within the tumor are observed. An increase in vp is observed in a few pixels. Over the following days these parameters tended to return to baseline. Per-patient HTML reports that depict all slices of raw qMRI images, processed qMRI images, and computed maps of ADC, T1, iAUC, ktrans, vp, and ve, from all scan dates of each subject, have been publicly shared (Supplementary Material B).
Median parameter changes in tumor are depicted in Figs. 2 and 3. Changes outside the Repeatability Coefficient (RC) range are also depicted, indicating significant differences with respect to baseline. A Bland-Altman plot of ADC measurement variability described by RC for 28 tumors from 18 patients with available ADC replicates at baseline is shown in Fig. 2a. For the remaining 3 (= 31−28) tumors, we applied the whole data RC = 0.8×10−3mm2/s. In the 31 tumors that were analyzable, a decrease in ADC on day 1 relative to baseline was observed in 24 tumors: 8 tumors were below −RC, 2 tumors above + RC, and the remaining did not have a significant change. In the next visit, ADC of 6 out of 16 tumors were below −RC and of 1 tumor was above + RC. In the last visit, ADC of two and one out of 9 tumors were below −RC and above + RC respectively (Fig. 2b).
On day 1 relative to baseline, tumor T1 decreased below –RC in 13 out of 24 tumors, and increased above + RC in 2 tumors. At the next visit, the corresponding numbers were 5 tumors below −RC and 4 tumors above + RC, out of 22 total evaluable tumors. In the last visit, T1 of 1 each of 7 tumors were below −RC and above + RC (Fig. 2c).
On Day 1 relative to baseline, 3 and 10 out of 26 tumors exhibited tumor iAUC changes that were below −RC and above + RC respectively. At the following visits, tumor iAUC of 8 and 3 out of 24 tumors, and 5 and 0 out of 6 tumors, were below −RC and above + RC (Fig. 2d).
Tumor ktrans of 0 and 9 tumors on day 1, 2 and 3 tumors on visit 2, and 4 and 0 tumors in the last visit, were below −RC and above + RC respectively relative to baseline (Fig. 3a).
Tumor ve of 2 and 10 tumors on day 1, 5 and 6 tumors on visit 2, and 2 and 0 tumors in the last visit, were below −RC and above + RC, respectively relative to baseline (Fig. 3b).
Tumor vp of 4 and 12 tumors on day 1, 7 and 11 tumors on visit 2, and 1 tumor each at the last visit, were below −RC and above + RC respectively relative to baseline (Fig. 3c).
The results in Figs. 1–3 suggest a response to PEGPH20 treatment that manifests as a decrease in ADC and T1 on day 1, possibly from a decrease in tumor water content due to HA depletion20. Results also show an increase in iAUC, ktrans, and vp on day 1, suggesting an increase in perfusion, permeability, and vascularity. ve also increased after treatment, suggesting a release of ECM space.
Correlations of median changes between parameters are shown in Fig. 4 (bottom row); whereas changes in ADC and T1 are significantly correlated to one another, changes in ve are correlated with changes in most other parameters.
Scatter plots of median changes vs. parameters at baseline are shown in Fig. 4 (top row). Most points in ∆ADC and ∆T1 are negative, confirming the tendency of ADC and T1 to decrease following PEGPH20 treatment. Also, there is a negative correlation, which suggests a dependency of response to baseline values. Significant changes in ADC and T1 below −RC are more likely in tumors with baseline ADC values above 1.46 × 10−3mm2/s and baseline T1 values above 0.54 s. Balanced Accuracy using these thresholds for both ADC and T1 are 72% and 82%.
Most changes in the other parameters are positive on day 1 (Fig. 4). For iAUC, there is a negative correlation to baseline values with most significant changes above + RC happening in tumors with baseline values below 9.2 mM-s (BA = 76%). Changes in ve and vp are also negatively correlated to baseline, with thresholds to separate significant and non-significant changes above + RC at 0.17 (BA = 68%) and 0.02 (BA = 60%) respectively. Changes in ktrans were non-significantly negatively correlated to baseline values, with an optimal threshold of 0.07 min− 1 (BA = 72%, p > 0.05).
No clear relation was observed between changes in quantitative MRI parameters and drug dose or drug pharmacokinetics (Cmax, Cmin, AUC). Also, no relation was observed with HA levels in stained tumors and survival (Tables 2–4 in Supplementary Material A).
To investigate the robustness of the results with respect to the manual annotations, an expert radiologist (JRC) re-annotated tumors in a subset of five patients on DW-MRI and DCE-MRI at each visit. Correlations of median tumor parameters between observers are shown in Fig. 5, revealing very high correlations (ICC ∼ 0.9).
Pixel-wise analysis
A pixel-wise analysis of all parameter values was carried out to develop a multivariable model for predicting post-PEGPH20 response in any parameter from baseline parameter values. Eleven patients have lesions imaged using DWI, T1, and DCE-MRI; however, in three patients either ADC, T1 or DCE parameters could not be computed at baseline; and in two patients (102-002-115, 102-002-126) the sequences were not all acquired in the same view. In all, it was possible to co-register DWI, T1, and DCE-MRI sequences across scan dates in six patients, as shown for one patient in Fig. 1, and these were used in the pixel-wise analysis. Parameter changes in pixels from these six patients between Day 1 and baseline were visualized using Principal Component Analysis (PCA) to identify two clusters (Fig. 2 in Supplementary Material A). As in the VOI analysis, each tumor pixel was considered to either be a pharmacodynamic responder (p-responder) or non-responder (p-non-responder) based on its parameter changes relative to its RC.
For ADC we applied the pixelwise RC per tumor when baseline repeats were available, or otherwise the group pixel-wise RC = 1.4 × 10− 3mm2/s obtained from all patients using equal numbers of pixels per patient. For the other parameters we used the same RC as in the VOI analysis. A pixel with a parameter change between Day 1 and baseline that is below −RC for ADC and T1, or above + RC for the other parameters, was labeled a “p-responder” pixel, and otherwise a “p-non-responder” pixel.
Using the pixelwise baseline parameter values as independent variables, and the pixel class (p-responder or p-non-responder) as the dependent variable, we trained a decision tree model using the machine learning software Weka (www.cms.waikato.ac.nz/ml/weka/). To mitigate model over-fitting, we separated pixels into two sets, one for training and one for validation. To avoid biasing results towards larger tumors with more pixels, during the training process we sampled an equal number of pixels per patient, equivalent to 85% pixels from the smallest tumor, with the remaining pixels used for validation. Additionally, to avoid problems with unbalanced data, the sampling process during model training was randomized to select equal numbers of “p-responder” and “p-non-responder” in the training set. After each training iteration, the resulting decision tree was applied to the validation set, and a BA per patient was obtained, and the process was repeated multiple times with random sampling from the six patients. After this process, the optimal model was a simple rule: pixels with baseline ve < 0.39 are predicted to be “p-responder”, and otherwise predicted to be “p-non-responder” (average BA = 74.4% in the validation set, average training BA = 69.2%). As each patient contributed data to the training and validation sets, we performed a leave-one-patient-out cross-validation to evaluate this simple model, obtaining an evaluation average BA = 65.6% across patients (Table 5 in Supplementary Material A).