Radiation and Risk of Liver, Biliary Tract, and Pancreatic Cancers among Atomic Bomb Survivors in Hiroshima and Nagasaki: 1958–2009

Although incidence rates of liver cancer are decreasing in areas of the world with high and intermediate rates, liver cancer remains the fourth-most common cause of cancer death globally (1 ) and the sixth-most common in Japan (2 ). One of the most common risk factors for liver cancer is chronic infection with hepatitis virus (3 ), the prevalence of which varies by geographical region (4, 5 ). In Japan, the incidence of liver cancer increased from the 1970s to the mid-1990s, reportedly due to an increased prevalence of hepatitis C virus (HCV) infection, particularly among individuals who were born between the mid-1920s and mid-1930s (6, 7 ). There is sufficient evidence that high-linear energy transfer (LET) radiation, such as plutonium (8 ) and thorium (9, 10 ), causes liver cancer. However, there is much less evidence that exposure to low-LET radiation causes liver cancer (11 ), even in a large-scale international pooled analysis of nuclear worker cohorts (12 ). To our knowledge, the Life Span Study (LSS) of atomic bomb survivors in Hiroshima and Nagasaki is the only study that has consistently reported a significantly increased risk of liver cancer associated with exposure to low-LET radiation (13–17 ). Because liver cancer is largely linked to chronic infection with HCV in Japan (18 ), the possibility of confounding and effect modification by HCV has been investigated extensively in studies of the atomic bomb survivors (19, 20 ).

Pancreatic and biliary tract cancers are often diagnosed at advanced stages when there is little chance of curative treatment. There are established risk factors (21, 22 ) (e.g., smoking and obesity for pancreatic cancer, gallstones and anomalies of the pancreaticobiliary duct for biliary tract cancer), but whether radiation exposure is associated with these cancers is unclear (11 ). To date, no evidence of a radiation-related risk of pancreatic or biliary tract cancer has been found in the LSS. Several other studies have reported increased risks of pancreatic cancer among those who received radiotherapy for benign or malignant diseases, although a dose-response relationship has not been established (11 ). Recently, however, some published studies of cancer survivors have indicated a dose-response relationship between doses from radiotherapy and risk of pancreatic cancer (23, 24 ).

The primary objective of the current study was to estimate radiation risks for liver, biliary tract and pancreatic cancers in the LSS. Our analyses extend the cohort follow-up 11 years beyond that used in the most recent LSS incidence report, make use of improved dose estimates and account for the effects of alcohol consumption, smoking and body mass index (BMI) in radiation risk assessment.

This study was conducted as part of an ongoing series of analyses regarding radiation risks for solid cancer in aggregate (25 ) and at major anatomical sites (26–29 ). The study design and eligibility criteria were identical to those for the all-solid-cancer analysis (25 ). In brief, potentially eligible participants were members of the LSS cohort (n = 120,321): survivors of the atomic bombings who were within 10 km of the hypocenter in Hiroshima or Nagasaki (n = 93,741) and city residents who were not in either city (not-in-city; NIC) at the time of the bombings (n = 26,580). We excluded participants who died between 1950 and 1958 or had a documented history of cancer before the start of cancer incidence follow-up (n = 8,317), who could not be traced using koseki (family registry) records (n = 86), and one person who was enrolled in duplicate. We also excluded participants whose dose of atomic bomb radiation was unknown (n = 6,473). The resulting number of eligible participants was 105,444.

Cancer incidence follow-up of the LSS began on January 1, 1958, when the cancer registries in Hiroshima and Nagasaki were established. Follow-up continued until December 31, 2009 for the current analysis. Vital status and causes of death were confirmed from nationwide records of koseki and death certificates. Information about incident cancer cases was obtained through linkage with the cancer registries in Hiroshima and Nagasaki, supplemented by Adult Health Study (AHS) health examinations and ABCC-RERF pathology programs. Cancer cases were classified using International Classification of Diseases for Oncology, Third Edition codes (ICD-O-3). The outcomes of interest were first primary malignant tumors of the liver, including liver (ICD-O-3 topography: C22.0) or intrahepatic bile ducts (IHBD) (C22.1); biliary tract, including gallbladder (C23) or other and unspecified parts of biliary tract (C24); and pancreas (C25). Censoring events were: a first primary cancer or death due to a cause other than liver, biliary tract or pancreatic cancer; loss to follow-up due to emigration; end of follow-up when the participants reached 110 years of age; and the end of the study period.

Eligible cases were first primary liver, biliary tract or pancreatic cancers diagnosed within the Hiroshima and Nagasaki cancer registry catchment areas during the follow-up period. Person-years were adjusted for expected migration out of the Hiroshima and Nagasaki catchment areas, with estimated migration rates stratified by city, sex, birth year and calendar year. Cases diagnosed only by autopsy under the ABCC-RERF program (autopsy-only) were not included as eligible cases and were censored at death; the numbers of cases were 46, 20 and 15 for liver, biliary tract and pancreatic cancers, respectively. Because such autopsies were intensively performed largely among survivors with higher radiation doses and only during the early years of follow-up, inclusion of autopsy-only cases could result in distorted radiation risk estimates among those who were older at the time of the bombings (25 ).

Dosimetry System 2002 (DS02) was used to estimate revised DS02R1 radiation doses (30, 31 ), which were adjusted for implausibly high dose estimates (>4 Gy) and random measurement errors (32 ). DS02R1 weighted absorbed liver doses, which were the sum of gamma ray and 10 times neutron doses, were used in all analyses of liver and biliary tract cancers. Weighted absorbed pancreas doses were used in analyses of pancreatic cancer.

Information regarding smoking, alcohol consumption, body height and body weight was collected through a series of LSS mail surveys (1965, 1969, 1978, 1991) and AHS questionnaires (1963, 1965, 1968) (33 ). We examined the consistency of responses across time and created summary variables for smoking (non-smoker, current smoker, former smoker or unknown) and drinking (non-drinker, current drinker, former drinker or unknown) status. We also calculated pack-years of cigarette smoking as cumulative smoking intensity. Initial pack-years was total pack-years at the time of the first questionnaire and pack-years accumulated as follow-up proceeded. Change in smoking status was incorporated into the computed pack-years by using information about age at starting and quitting smoking. Amount of alcohol intake (grams per day) was averaged over periods when a person was a drinker. For former drinkers, intake was assigned the value of their former intake amount. We calculated BMI as the weight in kilograms divided by the square of the height in meters using the earliest height and weight.

Case counts and accrued person-years of follow-up were stratified by city, sex, age at exposure, attained age, calendar time, NIC status, DS02R1 weighted absorbed liver (or pancreas) dose and an indicator of total shielded kerma >4 Gy. We used Poisson regression methods to estimate radiation risks of liver, biliary tract and pancreatic cancers. The risks were presented as the excess relative risk (ERR) and excess absolute risk (EAR) using:

where λ₀ represents the baseline (or background) rate for individuals without radiation exposure, and ERR_rad and EAR_rad represent the ERR and EAR for radiation exposure, respectively. EARs were estimated only when ERRs were statistically significant.

Exploratory data analysis, along with known trends in cancer in Japan, was used to inform modeling strategies. Sex-specific baseline rates were modeled using birth cohort categories (1914 or earlier, 1915–1924, 1925–1934, 1935–1945) for liver cancer or linear birth cohort effect for biliary tract and pancreatic cancers, and quadratic regression splines for log-attained age with a knot at age 70 years. For liver cancer, due to heterogeneity within and between cities in liver cancer rates by distance from the hypocenter (17 ), the model included an interaction between city and NIC status and an interaction between city and distal location (i.e., ≥3 km from the hypocenter), such that only zero-dose proximal (i.e., <3 km from the hypocenter) survivors were included in the reference group (34 ). For biliary tract and pancreatic cancers, because there was no evidence of heterogeneity by location, all zero-dose survivors and NIC residents were included in the reference group. However, the background models for biliary tract and pancreatic cancers included an interaction between sex and city to describe differences in incidence rates between men and women in Hiroshima and Nagasaki.

The shape of the radiation dose response was determined by comparing the fit of linear versus linear-quadratic dose-response models, for which comparisons were made using likelihood ratio tests and the Akaike information criterion (AIC). Curvature was defined as the ratio of the linear to quadratic dose-response coefficients. For all analyses, an indicator variable for total shielded kerma >4 Gy was included as an effect modifier. For liver cancer, effect modification of the ERR and EAR by sex, age at exposure and attained age was evaluated. The ERR and EAR were therefore presented as sex-averaged risks among individuals who were exposed at an age of 30 years and who attained an age of 70 years. We also computed radiation ERRs of liver cancer by categories of age at exposure to determine whether there was non-monotonic effect modification by age at exposure, as was observed in previously published studies (15–17 ). For biliary tract and pancreatic cancers, we considered effect modification on the ERR and EAR by sex, but lack of an overall dose response precluded consideration of effect modification by age at exposure and attained age in multivariable models. Instead, descriptive analyses compared crude incidence rates across dose groups, stratified by categories of age at exposure and attained age.

For liver and pancreatic cancers, joint effects of radiation with smoking, alcohol consumption and BMI were estimated using both additive and multiplicative ERR models and using EAR models (only for liver cancer). In the multiplicative ERR model, radiation risks are estimated relative to those who had the same smoking status, drinking status and BMI, whereas in the additive model, radiation risks are estimated relative to non-smoking non-drinkers with BMI <25.0 kg/ m². The fit of additive versus multiplicative models was compared using AIC. For smoking, the models included cumulative pack-years, with a corresponding ERR (ERR_smk). For liver cancer, sex-specific ERRs of smoking were estimated because of a better model fit. For alcohol consumption, the model included average intake amount, with a corresponding ERR (ERR_alc), as well as an indicator variable for former drinkers. Thus, the model allowed for risk differences between current and former drinkers with the same amount of (current or former) alcohol consumption. For BMI, the model included indicator variables for BMI ≥25.0 kg/m² (ERR_BMI) and unknown BMI, because the association between BMI and risk of liver or pancreatic cancers appeared to be J-shaped. For pancreatic cancer, sex-specific indicator variable for unknown BMI was used. Smoking and alcohol EARs (EAR_smk and EAR_alc) for liver cancer were computed using models with the same terms as those for ERR_smk and ERR_alc with effect modification by attained age. Indicator variables for BMI ≥25.0 kg/m² and unknown BMI were included in the baseline model in the EAR model.

We performed several sensitivity analyses for liver and biliary tract cancers. Because liver and IHBD cancers, as well as gallbladder and other and unspecified parts of biliary tract (e.g., extrahepatic bile duct, EHBD) cancers, differ with respect to their cell of origin, risk factors (35–37 ), and possibly radiation-induced carcinogenesis, we estimated radiation ERRs separately for liver and IHBD cancers and for gallbladder and EHBD cancers. The ERRs for liver, IHBD, gallbladder and EHBD cancers were calculated from ERR models fit to separate person-year datasets, as well as from joint analyses (38 ), which accommodated heterogeneity in background rates and radiation effects between liver and IHBD cancers, between gallbladder and EHBD cancers and between IHBD and EHBD cancers. First, person-year datasets for liver and IHBD cancers were combined, and the fitted model included interaction of an indicator variable for liver versus IHBD cancer with sex-specific attained age and birth cohort in baseline rate model, radiation dose, and effect modification by sex, age at exposure and attained age. Second, person-year datasets for gallbladder and EHBD cancers were combined, and the fitted model included interaction of an indicator variable for gallbladder versus EHBD cancer with sex-specific attained age and birth cohort in baseline rate model and radiation dose. Third, person-year datasets for IHBD and EHBD cancers were combined, and the fitted model included interaction of an indicator variable for IHBD versus EHBD cancer with sex-specific attained age and birth cohort in baseline rate model and radiation dose.

This study was approved by the RERF Institutional Ethical Review Board (research protocols 1-75 and 18-61). Use of data on causes of death and cancer incidence was approved by relevant authorities.

The number of eligible participants was 105,444, including 80,205 in-city survivors and 25,239 NIC residents.

During the follow-up period of 1958–2009, we ascertained 2,016 first primary liver cancer cases, comprised of 1,885 liver and 131 IHBD cancer cases, over 3,079,460-person years (Table 1); 522 cases (25.9%) were identified during the 11 years of follow-up since the last reported study. Dominant histological types were hepatocellular carcinoma and cholangiocarcinoma for liver and IHBD cancers, respectively (Table 2). The relatively low proportion of cases with histological verification and the high proportion of cases identified solely from death certificates reflects the relative inaccessibility to this organ and indicates poor diagnostic accuracy for this cancer.

The crude (unadjusted) liver cancer incidence rate was 6.5 per 10,000 person-years. The crude incidence rate among men (10.2 per 10,000 person-years) was more than twice that among women (4.4 per 10,000 person-years) (Table 1). Unadjusted rates exhibited a non-monotonic pattern across birth cohorts among men, with the 1925–1934 birth cohort having the highest rate, while the rates tended to decrease with increasing birth year among women. These patterns reflected birth-cohort-related differences in liver cancer incidence in the general Japanese population. Unadjusted rates increased with increasing attained age and radiation dose among both men and women. Rates among those exposed to <5 mGy in Hiroshima varied by location at the time of the bombing. In particular, the rate among distal survivors was higher than that among proximal survivors and NIC residents; the adjusted relative risk for distal survivors compared to proximal and NIC residents was 1.34 (95% CI: 1.17 to 1.52). In Nagasaki, there was no such difference.

There was no evidence of non-linearity in the radiation dose response (P = 0.344). When both the linear and quadratic dose-response terms were allowed to vary by sex, the curvature estimate was 0.05 for men (P > 0.50) and 2.16 for women (P = 0.180). Therefore, we focused on the linear dose-response model for both sexes. Sex-averaged ERRs per Gy estimated from the linear dose-response model with effect modification by sex, age at exposure and attained age are shown in Table 3 and Fig. 1. The sex-averaged ERR was significantly increased, with no significant difference between men and women (P = 0.371). There was no evidence of effect modification by age at exposure (P = 0.168) or attained age (P = 0.370). We also computed radiation ERRs among categories of age at exposure (Table 4). The ERRs were significantly increased among those who were 0–9 years (ERR per Gy 0.81), 10–19 years (ERR per Gy 0.66) and 20–29 years (ERR per Gy 0.92) at exposure, and decreased and became nonsignificant thereafter, although the difference in ERRs between age categories was not statistically significant (P = 0.378). There was also no evidence of non-linearity in the interaction between radiation and age at exposure on the log-linear scale: addition of a quadratic term for the age-at-exposure effect modifier to a model with a linear term did not improve model fit (P = 0.287). Therefore, we selected a log-linear model for effect modification by age at exposure.

The sex-averaged EAR per 10,000 person-year Gy was 5.32 overall, 6.90 among men and 3.74 among women, but the difference between sexes was not statistically significant (P = 0.107) (Table 5). The EAR significantly decreased with increasing age at exposure (P = 0.011). EARs within categories of age at exposure exhibited a similar pattern to those for the ERR. The EARs increased among those who were 0–9 years (10.34, 95% CI: 2.31 to 23.57), 10–19 years (12.90, 95% CI: 5.90 to 21.70) and 20–29 years (10.47, 95% CI: 2.82 to 21.05) at exposure, but decreased after 30 years old. In contrast to the ERR, the EAR increased with increasing attained age (P < 0.001).

We modeled joint effects of radiation with smoking, alcohol consumption and BMI using both additive and multiplicative ERR and EAR models. Although the fit of the additive ERR model was better than that of the multiplicative ERR model (AIC difference of 15.8), we preferred the multiplicative ERR model because of its simpler interpretation. The radiation ERR was essentially unchanged after adjustment for smoking, alcohol consumption and BMI (Table 3). These lifestyle factors were associated with liver cancer risk. The smoking ERR was 0.36 per 50 pack-years (95% CI: 0.16 to 0.60) among men and 1.20 per 50 pack-years (95% CI: 0.54 to 2.01) among women. The alcohol ERR was 0.09 per 20 g of ethanol intake per day (95% CI: 0.03 to 0.17). The ERR for former drinkers was significantly elevated (0.86, 95% CI: 0.52 to 1.25), possibly due to reverse causality: drinkers ceased their alcohol consumption due to health problems related to liver cancer. The ERRs for BMI ≥25 kg/m² were 0.46 (95% CI: 0.27 to 0.67). As noted with the ERR, adjustment for smoking and alcohol consumption did not affect the estimated EAR (Table 5).

There were 107 (75.0 plus 32.2; see Table 6) excess cases attributable to exposures to atomic bomb radiation including interaction with smoking, alcohol consumption and BMI (Table 6). The attributable fraction for those with ≥5 mGy of radiation was 11.6%.

Site-specific ERRs for liver and IHBD cancers were calculated (Supplementary Table S1; http://dx.doi.org/10.1667/RR15341.1.S1). The ERRs per Gy estimated from separate models (without effect modification) were 0.57 for liver and 0.70 for IHBD cancers. Similar ERRs per Gy were estimated by the joint model (without effect modification), and there was no significant difference between the ERRs for liver and IHBD cancer (P > 0.50). The inclusion of effect modifiers had little impact on the radiation ERRs obtained from separate and joint analyses.

There were 694 incident biliary tract cancer cases including 354 gallbladder and 340 other and unspecified parts of biliary tract (EHBD) cancers. The crude incidence rate was 2.3 per 10,000 person-years and was 1.9 among men and 2.4 among women (Table 7). Differences in fitted baseline rates between men and women differed between cities (P < 0.001); in Hiroshima, the adjusted relative risk between men and women was 0.70 (95% CI: 0.56 to 0.88), but in Nagasaki, the adjusted relative risk was 1.31 (95% CI: 0.73 to 2.36). The ERR per Gy estimated from a model without effect modification was –0.02. Sex-specific ERRs were 0.17 among men and –0.14 among women (Table 8), with no evidence of difference between sexes (P = 0.284). Crude incidence rate ratios by dose group did not differ across strata of age at exposure or attained age.

The ERR for gallbladder cancer was negative, while the ERR for EHBD cancer was positive (Supplementary Table S2; http://dx.doi.org/10.1667/RR15341.1.S1). However, neither were statistically significant, and the difference between the ERRs was not statistically significant (P = 0.183). The ERRs for IHBD and EHBD were both positive (Supplementary Table S2), but the difference between ERRs was not statistically significant (P = 0.379).

We ascertained 723 incident cases of pancreatic cancer (Table 9), of which 219 cases (30.3%) occurred during the 11 years of additional follow-up. Accuracy of diagnosis for pancreatic cancer was considered poor as suggested by the relatively high proportion of cases reported only by death certificate and the low proportion of cases diagnosed by histology (Table 2). The crude incidence rate of pancreatic cancer was 2.3 per 10,000 person-years and was 2.7 among men and 2.2 among women. Differences in fitted baseline rates between men and women differed between cities (P = 0.065); in Hiroshima, the adjusted relative risk between men and women was 0.48 (95% CI: 0.38 to 0.59), while in Nagasaki, the adjusted relative risk was 0.65 (95% CI: 0.38 to 1.15). Radiation risk estimates of pancreatic cancer are shown in Table 10. The sex-averaged ERR per Gy estimated by a model without effect modification by age at exposure or attained age was 0.38. Although the sex difference was not statistically significant (P = 0.193), women had a numerically higher ERR compared to men. Among women, the ERR for pancreatic cancer was significantly increased. Crude incidence rate ratios by dose group did not differ across strata of age at exposure or attained age. A model with multiplicative joint effects between radiation, smoking, alcohol consumption and BMI revealed a significantly increased sex-averaged ERR (0.45; 95% CI: 0.07 to 0.92). In addition, the ERR was significantly elevated among women, but not among men. The ERR for smoking was 0.83 per 50 pack-years (95% CI: 0.37 to 1.44), but the ERRs for alcohol consumption and BMI ≥ 25.0 kg/m² were not significantly increased.

The updated liver cancer incidence data in the LSS continued to provide evidence of a linear relationship between radiation dose and liver cancer risk, regardless of whether joint effects with smoking, alcohol consumption and BMI were considered. We analyzed liver and IHBD cancers combined and separately, and found a suggestion of an excess risk of IHBD cancer, although there was no evidence that radiation ERRs differed between liver and IHBD cancers. We also tested for differences in radiation risks between gallbladder and EHBD cancers, as well as between IHBD and EHBD cancers, and again there was no evidence of a difference. Analyses considering IHBD and EHBD, which were grouped with liver and gallbladder, respectively, were not typically performed in the LSS, with one exception (17 ). The extended follow-up increased the number of cases of biliary tract cancer (23.6%), but there was no evidence of radiation-associated excess risk. Pancreatic cancer (30.3%) also increased during the 11-year extended follow-up period, and there was a suggestion of excess risk among women.

In the current study, the estimated ERR (0.53 per Gy) and EAR (5.32 per 10,000 PY per Gy) for liver cancer, unadjusted for smoking, alcohol consumption and BMI, were higher than those in the previously published study by Preston et al. (ERR per Gy = 0.30 and EAR per 10,000 PY per Gy = 4.3) (16 ). We investigated whether revised doses (31 ), extension of the follow-up period or change in statistical models explained the difference by comparing ERRs estimated based on both previous and current data and statistical models. Both extension of the follow-up and change in statistical model (described below) increased the estimated ERRs, but the revised doses did not have a substantial effect (data not shown). In the current study, we selected zero-dose proximal survivors (<3 km from the hypocenter) as the reference group, while Preston et al., selected all zero-dose survivors (<10 km from the hypocenter) as the reference group. Because the <5 mGy proximal survivors had a lower incidence of liver cancer compared to the <5 mGy distal survivors (3 to <10 km from the hypocenter), perhaps due to urban-rural differences in socioeconomic or lifestyle factors associated with liver cancer, change in reference group could have resulted in an increased ERR (34 ). In fact, after fitting a model with all zero-dose survivors as the reference group, the ERR decreased from 0.53 (95% CI: 0.23 to 0.89) to 0.39 (95% CI: 0.14 to 0.70).

Previous LSS liver cancer incidence studies reported a non-monotonic age at exposure effect (Table 4). The ERRs were shown to be higher among those exposed between 10 and 39 years, peaking among those exposed between 10 and 19 years, with no or small risk among the youngest and the oldest age-at-exposure groups (15–17 ). In the current study, both absolute and relative risks by categorical ages at exposure were significantly increased for exposure under age 9 years, 10–19 years and 20–29 years, with no significant risks after age 30 years (Table 4). However, these risk differences between groups were not statistically significant. In the current study, the baseline liver cancer rates were relatively low among those who were 0–9 years at exposure (Table 1). This lower baseline rate might reflect a lower prevalence of HCV infection among this birth cohort, which was observed among general populations in Japan (7 ) as well as among the Adult Health Study (AHS) participants, a clinical subset of the LSS, in Hiroshima (39 ). Additional follow-up of this youngest birth cohort might provide useful information on how the prevalence of HCV infection influences radiation risk estimates of liver cancer.

Excess risks of radiation-associated liver cancer have been consistently reported in the LSS. However, outside the LSS cohort studies, there has been no clear evidence of excess liver cancer risks from populations with medical (40, 41 ) or occupational (12, 42 ) exposures to low-LET radiation. One of the few studies that found excess risk of radiation-associated liver cancer is a cohort study of physicians in China (43 ). Radiologists had elevated risk of developing incident liver cancer compared to physicians not engaged in radiological procedures. Nevertheless, a dose-response relationship was not demonstrated. In addition, radiation risk estimates could have been influenced by a higher prevalence of hepatitis virus infection among those who started radiological practice in earlier periods, as well as by diagnostic inaccuracies for primary liver cancer, which was also noted in a study in high natural background radiation areas in China (44 ).

In published studies on the effects of high-LET radiation such as thorium or plutonium, excess risks of liver cancer have been demonstrated. Patients who were chronically exposed to alpha particles from Thorotrast (contrast material containing colloidal thorium dioxide) (9, 10 ) and workers at the Mayak nuclear facility who were exposed to both inhaled plutonium and external gamma radiation (8, 45, 46 ) had elevated risks for liver cancer. Of note, the dominant histological type is hepatocellular carcinoma among the LSS participants (Table 2) and Mayak workers with relatively lower plutonium doses (46 ), while it is cholangiocarcinoma and/or hemangiosarcoma among Thorotrast patients, reflecting the fact that intravenously administered Thorostast is more likely to deposit in periportal areas than in hepatic cord areas of the liver (47 ). In the Mayak workers cohort, hemangiosarcoma was also observed, exclusively among workers exposed to higher plutonium doses (≥4 Gy) (46 ).

In the current study, radiation risk of pancreatic cancer was significantly increased among women, but not among men. Whether there is a sex difference in radiation risks of pancreatic cancer remains unknown. Studies among female patients who received radiotherapy for uterine bleeding (48 ) or cervical cancer (49 ) reported excess risks of pancreatic cancer associated with radiotherapy, but studies including both sexes (23, 41, 50 ) did not assess sex difference in risks. Although significant dose-response relationship was rarely demonstrated by earlier studies, more recently published studies reported that there was a significant dose-response relationship between doses from radiotherapy and risk of pancreatic cancer among survivors of Hodgkin lymphoma (23 ) or testicular cancer (24 ) while accounting for effects of chemotherapy. Reported excess odds ratio per Gy (0.12; 95% CI 0.03 to 0.42) among 5-year survivors of testicular cancer (24 ) was comparable with the ERR per Gy among male LSS participants (0.07; 95% CI –0.29 to 0.63). Studies among patients who received Thorotrast suggested that exposure to thorium might be associated with increased risks for pancreatic and gallbladder cancers (10 ), but there is no clear evidence of increased risk in other populations exposed to high-LET radiation. More studies with precise dosimetry and outcome information are required to further elucidate radiation risks of pancreatic or biliary tract cancers.

Although we did not have data on hepatitis virus infection, how infection with hepatitis virus affects radiation risk estimates of liver cancer in atomic bomb survivors is important, because Japan is an area where the prevalence of HCV infection is high and large proportions of liver cancer cases are attributable to HCV (3, 18 ). To date, there has been no indication of confounding by hepatitis virus infection. Studies in the AHS found that radiation dose was not associated with HCV infection (39 ). Moreover, estimated radiation risks were essentially unchanged after adjustment for infection with HCV in the AHS (20 ). A radiation dose-related increase in prevalence of chronic HBV infection was observed (51–53 ). However, this was limited to atomic bomb survivors who received blood transfusions between 1945 and 1972 when screening for blood product was not performed in Japan (53 ). Adult-onset persistent infection with HBV was established perhaps due to impaired immune function caused by radiation exposure (53 ). This suggests that HBV infection might be a mediator of the association between exposure to atomic bomb radiation and occurrence of liver cancer. A possibility that HCV might synergistically work with radiation to increase excess risks of liver cancer was reported elsewhere in a study of the LSS (19 ), although a more recently published analysis in the AHS did not provide support for this study (20 ). The role of hepatitis virus infection in the etiology of radiation-associated liver cancer should be more definitively demonstrated in future studies by developing a biologically plausible mechanistic framework.

A strength of our study is the large sample size with a sufficient number of cases over a long follow-up period, between 1958 and 2009. Because the LSS participants are comprised of both sexes and a wide range of ages at exposure, we were able to evaluate effect modification by these factors. We adjusted for effects of major non-radiation risk factors such as smoking, alcohol consumption and BMI, and demonstrated that radiation risk estimates of liver cancer were not influenced by these factors.

We acknowledge the following limitations. Information about infections with HBV and HCV was not available, although, as mentioned above, there has been no convincing evidence that HBV and HCV could confound radiation risk estimates for liver cancer, according to previously reported studies among atomic bomb survivors. Also, we considered joint effects with major risk factors for cancer such as smoking, alcohol consumption and BMI, which were collected after exposures to atomic bomb radiation. It might be possible that survivors with higher doses were more likely to change lifestyle habits because of health concerns, although survivors generally are unaware of their radiation dose. In addition, we accounted for reported changes in smoking and alcohol consumption over time, but we were unable to account for changes between the last questionnaire survey in 1991 and the end of follow-up in 2009, when older survivors might have decreased their smoking and alcohol consumption. Because approximately 76% of all person-years were accumulated before 1990 and because it takes many years for changes in these exposure levels to translate into appreciable changes in cancer risk (26 ), we do not expect that misclassification of lifestyle habits in recent calendar years would have a meaningful impact on the results. Another limitation is that a larger proportion of liver and pancreatic cancer cases were ascertained only by death certificates (DCs). Sharp et al. found a substantial impact of misclassification of primary liver cancer in DCs on estimated baseline liver cancer rate (54 ). The extent to which the DC misclassification affects radiation dose-response analysis remains unclear.

In conclusion, this study provided consistent evidence of a dose-response relationship between atomic bomb radiation and risk of liver cancer. Further follow-up is essential to clarify age-at-exposure effect on radiation risks of liver cancer. Our findings continue to suggest the lack of excess radiation-related risks for biliary tract cancer. There was a suggestion that radiation risk of pancreatic cancer was increased among women.

Table S1. Site-specific ERRs of liver (C22.0) and intrahepatic bile duct (C22.1) cancers.

Table S2. Site-specific ERRs of gallbladder (C23), extrahepatic bile duct (C24) and intrahepatic bile duct (C22.1) cancers.

The Radiation Effects Research Foundation (RERF; Hiroshima and Nagasaki, Japan) is a public interest foundation funded by the Japanese Ministry of Health, Labour and Welfare (MHLW) and the U.S. Department of Energy (DOE). The research was also funded in part through DOE award no. DE-HS0000031 to the National Academy of Sciences and contract no. HHSN261201400009C through the U.S. National Cancer Institute (NCI), with additional support by the Division of Cancer Epidemiology and Genetics in the NCI Intramural Research Program. This publication was supported by RERF Research Protocols 1-75 and 18-61. The views of the authors do not necessarily reflect those of the two governments.