Early post-treatment MRI predicts long-term hepatocellular carcinoma response to radiation segmentectomy

In this study, we included consecutive patients with HCC who underwent radiation segmentectomy between April 2014 and January 2018 from the interventional radiology database (n = 323). The diagnosis of HCC and the indication for radiation segmentectomy were based on imaging criteria (based on LI-RADS) and multidisciplinary tumor board discussion. Exclusion criteria were as follows: patients who underwent prior LRT (n = 73), lack of imaging follow-up at 12 months (n = 63), lack of contrast-enhanced MRI prior to radiation segmentectomy (n = 47), patients who underwent retreatment of the same lesion within 12 months (n = 35), lack of early post-treatment MRI (n = 10), other diagnosis than HCC on post-treatment pathology from liver resection (n = 7), status post liver transplant (n = 3), failed radiation segmentectomy (n = 2), adjuvant immunotherapy (n = 1), insufficient baseline MRI image quality (n = 1), and lesion size smaller than 1 cm (n = 1). The final study population consisted of 80 patients (M/F 60/20, mean age ± standard deviation 67.7 ± 9.9 years) (Fig. 1, Table 1).

Radiation segmentectomy was defined as transarterial Y90 infusion limited to two or less hepatic segments (Couinaud classification) in one treatment session [26]. The procedure and dosage methodology have been described previously [26‐28]. In short, a planning angiogram was performed prior to radiation segmentectomy to identify tumor-feeding vessels and inject 99Technetium‐macroaggregated albumin (99Tc‐MAA) for shunt detection and quantification. Patients were treated with 20–30-μm-sized glass-based microspheres (TheraSphere, Boston Scientific International) to transarterially deliver Y90 to the lesion and surrounding segment. The injected activity (GBq) was dependent on lesion size and the lung shunt fraction with intended target lesion dosing > 190 Gy.

The 1^st follow-up MRI was performed at 1.5 T (n = 50) or 3 T (n = 30) from variable vendors using gadoxetate disodium (Eovist/Primovist n = 69, using 10 mL fixed dose) or an extracellular contrast agent (n = 11, 0.1 mg/kg body weight). The median time between radiation segmentectomy and the 1^st follow-up MRI was 47 days (IQR, 43.0–53.5 days). Breath-hold axial pre-contrast 3D T1-weighted-images (T1WI), contrast-enhanced T1WI obtained during the arterial (2 phases back to back using fixed timing or bolus tracking method), portal venous (60 s after contrast injection), delayed/transitional phase (3 min after contrast injection) and hepatobiliary phase (only with gadoxetate disodium, 10 and 20 min after contrast injection), axial/coronal single-shot T2WI, and axial in- and opposed-phase imaging were available in all patients. Diffusion-weighted images (DWI, using b50, 400 and 800) was available in all but 6 patients. Automatically generated subtracted images (post-contrast images – pre-contrast images) were available for all contrast-enhanced phases in all patients.

Three independent radiologists (reader 1, M.J.K., a radiologist with 4 years of post-training experience in abdominal imaging; reader 2, M.E.H., an abdominal MRI fellow with 1 year of post-training experience; reader 3, J.G., a radiologist with 2 years of post-training experience) independently reviewed all early MRI studies in random order (obtained with a median delay of 47 days post-treatment). All lesions were indicated on screen shots on baseline MRI (placed by the study coordinator, D.S.) to ensure assessment of the correct index lesions. Baseline MRI before radiation segmentectomy and the 1^st follow-up MRI was available for the readout. All readers were blinded to clinical data and outcome. Evaluation of response assessment for the 1^st follow-up was performed on mRECIST, LI-RADS TRA, and image subtraction, respectively, in a random order for each patient in the same session. The different response criteria are summarized in Fig. 1 and Supplemental Table 1 (Supplemental document). According to mRECIST, the response of each lesion was characterized as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD), and as LR-TR nonviable, LR-TR equivocal, or LR-TR viable according to LI-RADS TRA version 2018. The percentage of necrosis was rated in 10% increments between 0 and 100% based on arterial or portal venous phase enhancement on subtraction images (Fig. 2 and Supplemental Table 1). All readers were instructed and trained using mRECIST, LI-RADS TRA, and image subtraction on a dataset (n = 7) that was not included in the final analysis. Tumor size was measured on portal venous phase images. We did not assess perilesional contrast enhancement in this study. Additionally, the quality of image subtraction at the 1^st follow-up was rated on a 5-point Likert-scale (based on liver edge visualization): 1, extensive misregistration, nondiagnostic images; 2, severe misregistration, images degraded but interpretable; 3, moderate misregistration, acceptable image quality with mild effects on diagnostic quality; 4, minimal misregistration, good image quality with minor or no effects on diagnostic quality; or 5, no misregistration, excellent image quality without effects on diagnostic quality.

The reference standard was based on a consensus readout of two separate radiologists (DS, the study coordinator, and BT, with 17 years of post-training experience in abdominal imaging). Both readers were blinded to the readout results from readers 1–3. The reference standard was based on mRECIST and was categorized as CR, PR, SD, or PD, respectively. We chose mRECIST as the reference standard, because it is currently the most commonly used method to assess treatment response in HCC after LRT.

Continuous data are reported as mean ± standard deviation or median with interquartile range (IQR) in parenthesis as appropriate. Shapiro–Wilk tests were used to test for normal distribution. Categorical data are reported as frequencies with percentages in parenthesis. Inter-reader agreement between the three readers and the agreement between the 3 different response assessment methods (mRECIST, LI-RADS TRA, subtraction) was assessed using Fleiss kappa (κ) [29]. A κ of < 0 indicated poor; 0.01–0.20, slight; 0.21–0.40, fair; 0.41–0.60, moderate; 0.61–0.80, substantial; and > 0.80, almost perfect agreement [30]. The interval data from the response assessment using image subtraction (0–100%, 10% increments) were categorized into four categories (1, 0–30%; 2, 40–60%; 3, 70–90%; 4, 100%) before calculating Fleiss kappa. To evaluate agreement between the response assessment methods and the prediction of CR at 12 months, we created binary categories (CR group and non-CR group) for the response assessment at the 1^st follow-up. The CR group at 1^st follow-up MRI was defined as CR according to mRECIST, LR-TR nonviable according to LI-RADS TRA, and 100% necrosis on subtraction, respectively. The non-CR group was defined as PR, SD, or PD according to mRECIST, LR-TR equivocal or LR-TR viable according to LI-RADS TRA, and 0–90% necrosis on subtraction, respectively. Subsequently, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy with 95% confidence intervals (95% CI) were calculated for the 1^st follow-up MRI, using the consensus readout at 12 months as the reference standard. Differences in diagnostic accuracy between mRECIST, LI-RADS TRA, and subtraction were calculated with Cochrane Q tests for each reader. Statistical analysis was performed using dedicated software (IBM® SPSS® Statistics 26; SPSS® Inc.). All tests were two-tailed. A p value < 0.05 was considered statistically significant.

Sensitivity, specificity, and PPV were fair to excellent for all readers for mRECIST (range for readers 1–3: sensitivity, 0.770–0.811; specificity, 0.667–1, and PPV, 0.966–1), LI-RADS TRA (range for readers 1–3: 0.676–0.811, 0.833–1, and 0.983–1), and subtraction (range for readers 1–3: 0.783–0.811, 0.667–1, and 0.967–1) when the 1^st follow-up MRI was used to predict CR at 12 months. NPV was low for all readers and methods (mRECIST, range 0.190–0.300; LI-RADS TRA, range 0.200–0.300; subtraction, range 0.200–0.300). No differences between response assessment methods regarding sensitivity, specificity, accuracy, PPV, and NPV were seen for reader 1 and reader 3. For reader 2, mRECIST and subtraction showed slightly higher sensitivity (0.770 and 0.783) and lower specificity (0.667 and 0.667) compared with LI-RADS TRA (sensitivity, 0.676; specificity, 1). No differences regarding diagnostic accuracy between mRECIST, LI-RADS TRA, and subtraction were found for reader 1 (p > 0.9), reader 2 (p = 0.053), and reader 3 (p > 0.9). Sensitivity, specificity, accuracy, PPV, and NPV for the prediction of CR at 12 months for all readers and criteria are presented in Table 2. An example for agreement and lack of agreement between the first follow-up and the 12-month follow-up is presented in Figs. 3, 4, and 5, respectively.

Readout results for patients with discordant response assessment between the 1^st follow-up and the 12-month follow-up are provided in Table 3. Patients with viable tumors at the 1^st follow-up but CR at the 12-month follow-up were most often rated as PR (reader 1: PR n = 12, SD n = 2; reader 2: PR n = 14, SD n = 2, PD n = 1; reader 3: PR n = 13, SD n = 2), viable (reader 1: viable n = 14; reader 2: viable n = 16, equivocal n = 8; reader 3: viable n = 14), and not 100% necrotic (reader 1: n = 14; reader 2: n = 16; reader 3: n = 15), respectively.

The inter-reader agreement between readers 1–3 was substantial for mRECIST (κ, 0.71; 95% CI, 0.54–0.90), LI-RADS TRA (κ, 0.62; 95% CI, 0.54–0.79), and subtraction (κ, 0.65; 95% CI 0.45–0.84). Our results suggest that experience might play a role in the response assessment with slightly higher diagnostic accuracy to predict CR for the most experienced radiologist (reader 1 with 4 years of post-training experience, 0.825), compared to the less experienced radiologist (reader 2 with 1 year of post-training experience, 0.700–0.775; reader 3 with 2 years of post-training experience, 0.800). Furthermore, the inter-reader agreement was higher for the two more experienced readers (reader 1 and 3) compared with the least experienced reader (reader 2).

The agreement between mRECIST, LI-RADS TRA, and subtraction for the 1^st follow-up was almost perfect for readers 1 and 3 (both κ, 1.0; 95% CI 0.64–1) and substantial for reader 2 (κ, 0.80; 95% CI, 0.64–0.97).

The majority of subtracted images were rated as good or excellent quality and none as not diagnostic (Supplementary Table 2).