Breast-iRRISC: a novel model for predicting the individualised lifetime risk of radiation-induced breast cancer from a single screening event

Abstract

Objectives:

This work establishes the prototype of a new innovative risk model that aims to evaluate the total risk involved with screening mammography for each individual female. This has been specifically designed to accommodate any combination of lifetime screening regimes, using only the information gathered from a single mammographic examination.

Methods:

This model prototype was developed with the aid of a large dataset of images from the Cancer Institute New South Wales (CINSW) with over 30,000 images from over 7000 examinations. Each examination is derived from a separate female.

Results:

This prototype which we have called Breast Individualised Risk of Radiation-Induced Screening Cancer (Breast-iRRISC) is a novel tool for the assessment of the lifetime risk involved with screening mammography. The results demonstrate the applicability of this approach to the various screening regimes utilised around the globe, in addition to the personalised screening frequency patterns females have undergone and are likely to receive in the future.

Conclusions:

This unique tailored approach to risk assessment will further empower females and clinicians towards a more informed clinical decision process regarding future imaging pathways. It will also inform health policy decisions regarding alternate screening durations and intervals.

Advances in knowledge:

Breast-iRRISC is a novel tool that provides females, clinicians and health policymakers around the globe with the ability to quantify the lifetime risk of radiation-induced breast cancer from screening mammography on an individual level from a single exposure.

Introduction

The linear extrapolation of risk from high to low dose is frequently debated; however, the evidence presented thus far is inconclusive.¹ The linear no-threshold (LNT) model is therefore the standard approach in use today, and although it may either overestimate or underestimate the risk, the International Commission for Radiological Protection (ICRP) stated in 2007 that efforts to disprove the validity of the LNT model may be “beyond scientific resolution” given the frequent contradictory evidence and observations.²

Assuming there is no lower threshold dose in which radiation exposure does not cause damage, repeated exposure to ionising radiation from screening mammographic procedures may increase the risk of breast cancer incidence.³ The risk of radiation-induced breast cancer from digital mammography screening depends on the female’s breast size and composition, in addition to the exposure parameters set on the mammography unit.⁴ The average female in Australia receives a radiation dose of around 1.39 milligray (mGy) from a single digital mammography screening projection and an average total of 5.68 mGy per examination for the four standard projections.⁵

There have been previous efforts to quantify the risk associated with mammography screening. In 2011, the lifetime risk of radiation-induced breast cancer and resultant death was calculated to be 86 and 11 per 100,000 females, respectively, when screened annually from 40 to 55 years and biennially until 74 years. These risk calculations were performed for both a single mammographic examination and for a select number of screening regimens by Yaffe and Mainprize. They employed a set dose of 3.7 mGy for each examination to assess lifetime risk, in addition to providing these results for a 1-mGy examination dose, enabling scaling of the risk to different doses.⁶ Formulations in 2016 devised by Warren et al to calculate the total number of induced cancers for one year of screening, took into account the number of screening rounds and the total number of females screened in a year.⁷ Also in 2016, Miglioretti et al incorporated a larger variety of more traditional screening regimens (annual, biennial and a hybrid strategy) to assess lifetime risk, stratified by compressed breast thickness (CBT). This approach sourced the examination radiation dose from the digital mammographic imaging screening trial (DMIST) distribution, conditional on the female’s CBT, which if it was ≥ 7.5 cm (i.e. 8% of DMIST population) was assumed to require extra views for a complete examination.⁴ In 2018, conversion factors were proposed by M.Ali et al for the calculation of radiation-induced cancer risk for an annual, biennial and triennial mammography screening frequencies for a commencement age from 25 to 75 years, assuming a cessation age of 75 years.⁸ This was to refine the earlier published mathematical model that attempted to estimate lifetime risk from screening mammography.⁹

The above approaches have two key limitations: firstly, there are a finite number of screening strategies and frequencies employed by any one study that assume no deviation from that attendance pattern. For example M.Ali et al only provide annual, biennial and triennial screening strategies assuming constant attendance until a cessation age of 75 years,⁸ Yaffe and Mainprize only provide six different screening regimens that include various combinations of annual and biennial screening⁶ and Miglioretti et al provide eight different screening patterns of annual, biennial and hybrid strategies.⁴ Secondly, they either use a set mammographic dose to estimate risk for all ages, such as an average dose of 3.7 mGy and 2.019 mGy for both breasts per examination in Yaffe and Mainprize⁶ and M.Ali et al,⁸ respectively, or at best estimate the dose per view conditional on the female’s CBT as in Miglioretti et al.⁴ These limitations ultimately mean that the approaches are not woman-specific and therefore cater more to the ‘average female’.

This study therefore aims to introduce a model that can predict the individualised lifetime risk of radiation-induced breast cancer within screening mammography for any standardised or personalised screening regime. It uses the information from the Digital Imaging and Communications in Medicine (DICOM) header from any single mammographic examination such as the female’s age, mammographic breast density (MBD), CBT, mAs, kVp, along with the half value layer (HVL) and constants that are calculated using the physics quality assurance (QA) data to model that female’s risk over a lifetime of screening.

Methods

The proposed model aims to predict the lifetime risk of radiation-induced breast cancer in screening mammography, using the information gathered from a single mammographic examination. This was done by the retrospective analysis and modelling of a dataset of images obtained from the Cancer Institute New South Wales (CINSW) containing 31,097 mammograms from 7,728 separate examinations, to establish the foundations of the model.

Three key steps were required to calculate a female’s overall cumulative risk of radiation-induced breast cancer, for their lifetime screening attendance, using data derived from a single mammographic examination. These steps were:

1. Estimating MBD change with age

   1. Plotting the average MBD (MBD_A) as a function of age

   2. Using the MBD_A to calculate the woman-specific MBD (MBD_WS) across screening ages.

2. Translating MBD to mean glandular dose (MGD)

   1. Using the MBD_WS to estimate the woman-specific CBT (CBT_WS)

   2. Applying a correction factor (µ_CBT) to the CBT_WS

   3. Extracting the woman-specific conversion factors required for the MGD calculation

   4. Calculating the woman-specific MGD (MGD_WS) with the aid of previous steps across the screening ages.

3. Translating MGD to effective risk (R)

   1. Calculating the cancer lifetime attributable risk (LAR) for breast tissue

   2. Using the LAR and MGD_WS across all screening ages to calculate the lifetime effective risk.

STEP 1: Estimating MBD change with age

This first step was to estimate the average change in MBD with age. To do this, the females’ MBDs were extracted from the CINSW data-set and filtered to include only those females aged between 40 and 75 years, in line with mammographic screening ages recommended in Australia. This resulted in 30,562 mammograms from 7,596 examinations. The MBD was estimated using LIBRA software and is detailed by Suleiman et al.¹⁰

STEP 1a: Constructing an MBD_A vs age curve

The extracted MBDs were then plotted against each age between 40 and 75 years (Figure 1). Equation 1 was derived from Figure 1 and describes the change and average estimation of the females’ MBD as a function of age in the following quadratic form:

Figure 1.

Estimation of average MBD (MBD_A) as a function of age created from 30,562 mammograms (7,596 examinations) for females aged 40–75 years. The error bars represent the standard deviation and the trendline added is a second-order polynomial that illustrates the fall in MBD with age.

$${\text{MBD}}_{\text{A}}\text{=}{\text{0.0163A}}^{\text{2}}\text{–2.2795A+89.778}$$ (1)

where MBD_A is the average MBD as a function of age and A is age in years. This was used to estimate the change in the females’ MBD throughout ages relevant to the screening regime.

STEP 1b: Constructing a MBD_WS vs age curve

The MBD_A for each age described by Equation 1 was then used to calculate MBD_WS for each female being examined. This was calculated by dividing the MBD in the female’s presented mammogram, by the corresponding average MBD at that age. This was achieved using Equation 1 by substituting the female’s age at the time of the screening in A and solving for the MBD_A. The newly calculated woman-specific correction factor (µ_WS) was then applied to the average MBDs across all screening ages between 40 and 75 and plotted as a function of age.

The following example will demonstrate this step:

Female X has a scan taken at age 63, presenting with an MBD of 29.5% (i.e. average MBD from the four projections taken, Mediolateral Oblique (MLO) and Craniocaudal (CC) of each breast). This is then divided by the MBD_A of a female at that same age of 63 years, i.e. 10.86% MBD, which was calculated using Equation 1 from Figure 1. The ratio difference between the two MBDs are calculated:

29.5 ÷ 10.86 ≅ 2.72… (two dp)

This µ_WS is then applied to the MBD_A for all ages of screening between 40 and 75 years (Figure 2). The µ_WS calculated will be different for every female.

Figure 2.

Female X example: new woman-specific MBDs (MBD_WS) plotted as a function of age (open circles) for female X for screening ages 40–75 years. This was done by applying μ_WS to the MBD_A (full circles) for all screening ages.

Equation 2i was derived from Figure 2 and describes this process:

$${\displaystyle MB{D}_{ws}=MB{D}_A\times {\mu }_{ws}}$$ (2)

where MBD_WS is the new woman-specific MBDs. The µ_WS must be applied for each age at screening.

Note: the MBD_WS in Figure 2 are an example of a single sample female, with the only confirmed MBD at the age of 63 years with which they presented. All other years of screening from 40 to 75 years are an estimated calculation based on the custom correction factor µ_WS that is calculated for each female.

STEP 2: Translating MBD to MGD

The next step was to convert the newly acquired MBD_WS to MGD values. This was done by using the following equation from Dance et al¹¹ for each MBD_WS, i.e. for every age at screening from 40 to 75 years:

$$\text{MGD=Kgcs}$$ (3)

where MGD is mean glandular dose. K is the air kerma in mGy and is calculated using the inverse square law and the tube output information from the physics QA guidelines of the mammography machines used.^12,13 The g-, c- and s-factor are conversion factors. g-factor accounts for a 50:50 MBD breast model and depends on HVL and CBT, c-factor accounts for the female’s specific glandularity and depends on HVL, CBT and MBD, s-factor accounts for the different X-ray spectra used, i.e. the anode/filter combination used.

STEP 2a: Constructing a CBT vs MBD curve

To account for the change in a females’ CBT which is needed when addressing the g- and c-factors for each age at screening in Equation 3, we can use the already determined MBD_WS to predict the CBT_WS at that density. Figure 3 shows CBT plotted as a function of all MBDs from 1–99% for females aged between 40 and 89 years. Equation 4i was derived from Figure 3 and describes the average relationship between females’ CBT and MBD in the following cubic form:

Figure 3.

Average estimation of CBT for different MBD values ranging from 1 to 99%. Constructed from 31,097 mammograms from 7,728 examinations of females aged between 40 and 89 years. The error bars represent the standard deviation.

$$\begin{gathered} {\displaystyle {\text{CBT}}_{\text{WS}}\text{=}{\text{–0.0001(MBD}}_{\text{WS}}{\text{)}}^{\text{3}}\text{+}{\text{0.0186(MBD}}_{\text{WS}}{\text{)}}^{\text{2}}\text{–} \\ {\text{1.2231(MBD}}_{\text{WS}}\text{)+ 69.619}} \end{gathered}$$ (4)

where CBT_WS is the woman-specific CBT as a function of the MBD_WS.

STEP 2b: The CBT_WS correction factor

Due to the averaging limitations of the approach proposed in Figure 3, there needs to be a correction factor applied to the CBT_WS for all ages at screening. Similar to step 1b, this was done by dividing the presented CBT in the female’s image, by the corresponding CBT_WS at that age, which can be calculated using Equation 4 by substituting the female’s MBD_WS at the presented age. The newly calculated CBT correction factor (µ_CBT) was then applied to the CBT_WS across all screening ages between 40 and 75 years (Figure 4).

Figure 4.

Female Y example: corrected estimation of CBT_WS plotted as a function of all MBDs from 1 to 99% (open circles) for female Y. This was done by applying the µ_CBT to the CBT_WS (full circles) calculated earlier.

The following example will demonstrate this step:

Female Y has a scan taken at age 65, with an average MBD of 5.5% and an average CBT of 60.75 mm (averages are from the four projections taken, two on each breast). This original CBT is then divided by the CBT_WS of female Y at the same age of 65 years, i.e. 63.44 mm, which was calculated using Equation 4 from Figure 3. The ratio difference between the two CBTs are calculated, i.e. 60.75/63.44 ≅ 0.96. This correction factor (µ_CBT) is applied to the CBT_WS for all MBD_WS for all ages of screening between 40 and 75 years (Figure 4). The µ_CBT calculated will be different for every female.

Equation 5i was derived from Figure 4 and describes this process:

$${\displaystyle CB{T}_C=CB{T}_{ws}\times {\mu }_{CBT}}$$ (5)

where CBT_C is the new corrected CBT_WS. The µ_CBT must be applied for each corresponding MBD_WS that was calculated with Equation 2.

STEP 2c: Conversion factors

g-factor: the HVL from the DICOM header and the CBT_C from step 2b are used to extract the g-factors for each age of screening from a table provided by Dance et al.¹¹

c-factor: the HVL from the DICOM header, the CBT_C from step 2b and the MBD_WS from step 1b are used to extract the c-factors for each age of screening from a table provided by Dance et al.¹¹

s-factor: the anode/filter combination extracted from the DICOM header of the original scan is used across all calculations for that female.

STEP 2d: Calculating MGD

Using the factors detailed in the above steps, the MGD_WS was calculated for each age at screening using Equation 3. The MGD_WS for each image (generally four projections, 2x MLO and 2x CC) was then added together to calculate the total dose delivered at each screening visit. This was done for all ages of screening, resulting in a list of MGD_WS values.

STEP 3: Translating MGD to lifetime effective risk

The final step was to convert the MGD_WS to effective risk (R) values. This was done by using the following equation from Brenner¹⁴ for each MGD_WS calculated for each age at screening in step 3:

$${\displaystyle R=\sum r_T{H}_T}$$ (6)

where R is the effective risk, r_T is the cancer LAR for tissue T, and H_T is the dose for tissue T. Note: r_T and H_T must be in the same dose units.

The LAR of breast cancer incidence for a given dose were taken from the BEIR VII Phase 2 report.¹⁵ These values are summarised in Table 1 and graphed in Figure 5.

Table 1.

The lifetime attributable risk of cancer incidence for breast tissue (¹⁵)

Age at Exposure (Years)	20	30	40	50	60	70	80
Cancer Incidence per 100,000 per 1 mGy	4.29	2.53	1.41	0.7	0.31	0.12	0.04

Figure 5.

The lifetime attributable risk (LAR) of cancer incidence for breast tissue per 100,000 females exposed at a single dose of 1 mGy. Source: Health Risks from Exposure to Low Levels of Ionising Radiation: BEIR VII.

STEP 3a: Calculating the LAR for breast tissue

To calculate the r_T values in Equation 6 across age, we can interpolate the LAR values in Table 1 to determine the LAR of breast cancer incidence for females aged between 20 and 80 years using the following two cubic equations derived from Figure 5:

$${\mathit{L}\mathit{A}\mathit{R}}_{\mathit{I}}\text{=}{\text{–0.00002A}}^{\text{3}}\text{+}{\text{0.0047A}}^{\text{2}}\text{–0.3694A+9.9493}$$ (7)

where LAR_I is the lifetime attributable risk of breast cancer incidence in cases per 100,000 females for a single exposure of 1 mGy and A is age in years.

STEP 3b: Calculating lifetime effective risk

Using the list of MGD_WS values obtained in step 2 (H_T) and the corresponding age-appropriate LAR from Equation 7 (r_T), we can calculate R for each age at screening using Equation 6. Depending on the frequency with which a female has undergone a mammogram, the R across each age screened is summed (∑) to calculate the lifetime effective risk for that female.

To demonstrate the usefulness of this work, five anonymous examples of females (female A-E), covering the following range of dose values and therefore risk of radiation-induced breast cancer: A) minimum, B) low, C) medium, D) high, and E) maximum, and also five anonymous centres (centre A-E) were chosen, averaging 179, 160, 156, 169, and 161 participating females, respectively, to cover a range of screening practices. These females were then applied to the model to calculate the number of induced cancers per 100,000 females, assuming the screening frequency for a variety of national screening programs, namely Australia, the United Kingdom (UK) and the United States (US).

Results

The preliminary results achieved with this model are shown. The risk of radiation-induced breast cancer per 100,000 females, for the examples of females A to E and centres A to E are summarised in Tables 2 and 3, respectively. This was produced based on applying our model to the data on age, MBD and CBT, that is extracted from the DICOM header from one mammographic event.

Table 2.

The lifetime risk of radiation-induced breast cancer for five anonymous females (A-E) presented for attendance at the national breast screening programs of Australia, the United Kingdom or the United States.

	Presented Information	Screening Regime	Screening Frequency	No. of Examinations	Cancers Induced (cases/100,000)
Female A (min risk)	Age: 40 years MBD: 78.25% CBT: 25 mm	Australia	Annually from 40 to 49 years Biennial to 74 years	23	29.9
			Biennially from 50 to 74 years	13	9.7
		United Kingdom	Triennially from 50 to 70	9	6.8
		United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	33.1
Female B (low risk)	Age: 56 years MBD: 8% CBT: 52.5 mm	Australia	Annually from 40 to 49 years Biennial to 74 years	23	65.9
			Biennially from 50 to 74 years	13	20.4
		United Kingdom	Triennially from 50 to 70	9	14.2
		United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	72.6
Female C (med risk)	Age: 68 years MBD: 16.25% CBT: 47.5 mm	Australia	Annually from 40 to 49 years Biennial to 74 years	23	100.3
			Biennially from 50 to 74 years	13	29.6
		United Kingdom	Triennially from 50 to 70	9	20.8
		United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	110.6
Female D (high risk)	Age: 67 years MBD: 10.5% CBT: 67.25 mm	Australia	Annually from 40 to 49 years Biennial to 74 years	23	135.0
			Biennially from 50 to 74 years	13	39.9
		United Kingdom	Triennially from 50 to 70	9	28.1
		United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	149.5
Female E (max risk)	Age: 56 years MBD: 3.75% CBT: 83.5 mm	Australia	Annually from 40 to 49 years Biennial to 74 years	23	172.5
			Biennially from 50 to 74 years	13	52.4
		United Kingdom	Triennially from 50 to 70	9	36.7
		United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	189.9

Table 3.

The lifetime risk of radiation-induced breast cancer for five anonymous centres (A-E) presented for attendance at the national breast screening programs of Australia, the United Kingdom or the United States.

	Screening Regime	Screening Frequency	No. of Examination	Cancers Induced (cases/100,000)
Centre A (179 females)	Australia	Annually from 40 to 49 years Biennial to 74 years	23	88.6 (SD = 13.1)
		Biennially from 50 to 74 years	13	26.1 (SD = 3.9)
	United Kingdom	Triennially from 50 to 70	9	18.3 (SD = 2.7)
	United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	97.4 (SD = 14.4)
Centre B (160 females)	Australia	Annually from 40 to 49 years Biennial to 74 years	23	79.2 (SD = 13.6)
		Biennially from 50 to 74 years	13	23.5 (SD = 4.0)
	United Kingdom	Triennially from 50 to 70	9	16.5 (SD = 2.8)
	United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	87.1 (SD = 14.9)
Centre C (156 females)	Australia	Annually from 40 to 49 years Biennial to 74 years	23	90.0 (SD = 12.4)
		Biennially from 50 to 74 years	13	26.8 (SD = 3.6)
	United Kingdom	Triennially from 50 to 70	9	18.7 (SD = 2.5)
	United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	99.0 (SD = 13.6)
Centre D (169 females)	Australia	Annually from 40 to 49 years Biennial to 74 years	23	111.9 (SD = 30.7)
		Biennially from 50 to 74 years	13	32.9 (SD = 8.9)
	United Kingdom	Triennially from 50 to 70	9	23.1 (SD = 6.3)
	United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	123.0 (SD = 33.7)
Centre E (161 females)	Australia	Annually from 40 to 49 years Biennial to 74 years	23	117.2 (SD = 30.9)
		Biennially from 50 to 74 years	13	34.2 (SD = 8.8)
	United Kingdom	Triennially from 50 to 70	9	24.0 (SD = 6.2)
	United States	Annually from 45 to 54 years Biennially ≥ 55 years	25	128.9 (SD = 34.0)

As an example, if a female with medium risk (female C) was to attend biennial screening (Breast Screen Australia’s active invitation age) from 50 to 74 years, they would expect a risk of about 29.6 radiation-induced breast cancers per 100,000 females (Table 2). This is equivalent to about a 0.03% increased risk to the 1-in-8 chance (i.e. 12.5%) of developing breast cancer over an 80 year lifespan.¹⁶

Note: any screening frequency pattern can be applied to this model, i.e. taking into account the exact history and intended future screening pattern a female is likely to receive. The screening regimes displayed in Table 2 and Table 3 were chosen based on the national protocols employed in Australia, the UK and the US.

Discussion

The proposed model is a prototype developed to predict the lifetime risk of radiation-induced breast cancer generated from a screening life-cycle, using the information derived from a single mammographic examination. The preliminary results indicate a wide range of risks amongst individual females (Table 2) as a direct consequence of breast density and compressibility. The applicability of our approach to different countries regardless of the screening regimen employed is shown (Table 3).

There are a few limitations with the current approach. The LAR estimates in Table 1 were derived from the BEIR VII – Phase 2 report¹⁵ with a minor adjustment; the units of risk were changed from cases/100,000 per 0.1 Gy as reported to cases/100,000 per 1 mGy as in Table 1. Although risk is dose dependant and it cannot strictly be assumed that the risk is 100 times less with 100 times less dose, we have taken the LNT approach as recommended by the ICRP given the frequent contradictory evidence provided for the effects of radiation at low doses.

The estimation of a female’s MBD throughout their screening lifetime is calculated from the MBD_A using the woman-specific correction factor µ_WS. Therefore, all MBD_WS curves are derived from a single MBD_A curve, which can cause inaccuracies for extreme cases with very high MBD. As an example, in Figure 2 the ~70% density of female X at age 40 is only a calculation; in reality she may have a lower breast density at that age. This limitation can be addressed by having multiple MBD curves, each corresponding to a percentile density category, in an attempt to assign the most appropriate percentile distributions of breast density to a particular female, instead of relying on the single MBD_A curve.

The next stage will be to further train the model by applying our approach to a much larger dataset of images that follow females throughout their screening journey. This will address many of the current limitations that result from having disconnected, single exposure data for estimating the lifetime MBD and CBT.Once this is achieved, the model’s dose output predictions will be validated against a separate cohort of real-world data to assess for accuracy. To do this, a dataset consisting of over 45,000 examinations from over 8000 females who have each undergone between 5 to 6 rounds of mammography screening has been obtained for this purpose.

Conclusion

This model is currently in its early stages, however upon training on a larger dataset and independent validation it should empower females and clinicians enabling an enhanced informed consent discussion regarding the benefits and risks of the mammographic screening process. It will also inform health policy makers who are considering the possible introduction of alternate screening durations and intervals. The proposed model will ultimately be available as an on-line platform that will be accessible to all females, doctors and health policy makers free of charge.