The impact of using weight estimated from mammographic images vs self-reported weight on breast cancer risk calculation

Abstract

Personalised breast screening requires assessment of individual risk of breast cancer, of which one contributory factor is weight. Self-reported weight has been used for this purpose, but may be unreliable. We explore the use of volume of fat in the breast, measured from digital mammograms. Volumetric breast density measurements were used to determine the volume of fat in the breasts of 40,431 women taking part in the Predicting Risk Of Cancer At Screening (PROCAS) study. Tyrer-Cuzick risk using self-reported weight was calculated for each woman. Weight was also estimated from the relationship between self-reported weight and breast fat volume in the cohort, and used to re-calculate Tyrer-Cuzick risk. Women were assigned to risk categories according to 10 year risk (below average <2%, average 2-3.49%, above average 3.5-4.99%, moderate 5-7.99%, high ≥8%) and the original and re-calculated Tyrer-Cuzick risks were compared. Of the 716 women diagnosed with breast cancer during the study, 15 (2.1%) moved into a lower risk category, and 37 (5.2%) moved into a higher category when using weight estimated from breast fat volume. Of the 39,715 women without a cancer diagnosis, 1009 (2.5%) moved into a lower risk category, and 1721 (4.3%) into a higher risk category. The majority of changes were between below average and average risk categories (38.5% of those with a cancer diagnosis, and 34.6% of those without). No individual moved more than one risk group. Automated breast fat measures may provide a suitable alternative to self-reported weight for risk assessment in personalized screening.

Introduction

A ‘one-size-fits-all’ model of breast screening has been widely adopted worldwide, although there is now growing interest in personalization to take into account both individual risk of breast cancer and the efficacy of mammography as a screening tool^1,2. Estimation of individual risk can be achieved via risk models which take into account personal risk factors such as hormone replacement therapy (HRT) use, parity, family history of breast cancer and body mass index (BMI)^3,4. It is often impractical to gather such data during attendance at screening, due to lack of space and privacy on mobile facilities, and short appointment times which maximize screening throughput. Self-reported risk data provide a convenient alternative. It has previously been reported that self-reported weight is unreliable, particularly at the upper and lower ends of the spectrum⁵ so we are exploring whether composition measurements from a woman’s mammogram can be used to derive an estimate of weight suitable for use in risk models. Breast fat volume measured by Volpara^™ has been shown to provide the best estimate of a woman’s weight⁶.

Methods

Between 2009 and 2015, 57,904 women recruited to the Predicting Risk Of Cancer At Screening (PROCAS) study provided self-reported risk information including self-reported height and weight, hormonal details and family history to enable calculation of breast cancer risk. Breast density was measured from the mammograms using Volpara^™ 1.5.0 (Volpara Health Technologies, Wellington, New Zealand).

Those with implants, mastectomies, a previous cancer diagnosis, missing risk data, fewer than four images with mammographic density measured with Volpara, or unfeasible height and weight values (weight < 4 st or weight > 35 st, height < 4 ft 4 in or height > 7ft), were excluded from the analysis. Breast cancer risk was calculated for the remaining 40,431 women using self-reported data in the IBIS risk calculator (version 6.0) which employs the Tyrer-Cuzick risk model³. Risk categories were assigned as follows based on 10 year risk from the model: below average <2%, average 2-3.49%, above average 3.5-4.99%, moderate 5-7.99%, high ≥8%.

In order to estimate weight from the mammograms, breast fat volume was computed for each image by subtracting the fibroglandular volume from the total breast volume computed by Volpara. The average breast fat volume for each woman was calculated from the four mammographic projections (cranio-caudal and mediolateral oblique views of each breast). The relationship between these values and the self-reported weights was established by Pearson correlation and linear regression. From this relationship, an estimated weight was derived for each woman using the predicted values from the linear regression, and used to calculate a second estimate of the Tyrer-Cuzick risk score.

Women diagnosed with breast cancer at the time of screening or subsequently were examined separately to those women without a cancer diagnosis (716 and 39,715 women respectively). A weighted kappa statistic was used to assess the agreement in the risk categories based on self-reported and estimated weight. The weighted kappa statistic gives more importance to larger changes in risk groups than single category changes⁷.

Univariate logistic regression was employed to assess the relationship (Odds Ratio (OR) per standard deviation (SD)) between weight (using the natural logarithm) and risk of cancer in the cohort.

Results

The self-reported weights of women in the sample ranged from 34.9 kg to 185.0 kg, with a median of 69.85 kg. The breast fat volumes of the women ranged from 24.0 cm³ to 3922.2 cm³ with a median of 789.55 cm³.

Figure 1 shows the scatterplot of self-reported weight vs average fat volume, with a regression line superimposed. The Pearson correlation coefficient was 0.709 (p<0.05). Weight values were estimated based on the results from the linear regression and the range of estimated values was 53.4 kg to 136.9 kg, with a median of 69.80 kg.

Figure 1. Relationship between self-reported weight and average fat volume in cm³ (Volpara).

In Table 1 the number of women within each risk category is shown for the two methods, in addition to the proportion of cancer cases ?within each risk category.

Table 1. Number of women (and percentage of cancer cases) by risk category using self-reported and estimated weights.

	Self reported			Estimated
	Cancer	Non-cancer	%	Cancer	Non-cancer	%
Below average	139	7146	1.91	125	6578	1.86
Average	415	22626	1.80	424	23132	1.80
Above average	99	5971	1.63	100	5977	1.65
Moderate	55	|3464	1.56	60	3494	1.69
High	8	508	1.55	7	534	1.29
Total	716	39715	1.77	716	39715	1.77

Tables 2 and 3 illustrate the impact of using estimated weights on risk category in women who had cancer at the time of screening or a subsequent cancer diagnosis (Table 2) and in those women who did not have a diagnosis of breast cancer (Table 3) separately.

Table 2.

Number and percentage of women with cancer diagnosed at the time of screening or subsequently by risk category using self- reported and estimated weights. Cells shaded green represent women for whom the risk category remained unchanged when using estimated weight; those shaded pink indicate a single category increase and those shaded blue indicate a single category decrease using estimated weight.

		Risk using estimated weight
		Below Average	Average	Above Average	Moderate	High	n (%)
Risk using self-reported weight	Below Average	119(16.6)	20 (2.8)	0	0	0	139 (19.4)
	Average	6 (0.8)	398 (55.6)	11(1.5)	0	0	415 (58.0)
	Above Average	0	6 (0.8)	87 (12.2)	6 (0.8)	0	99(13.8)
	Moderate	0	0	2 (0.3)	53 (7.4)	0	55 (7.7)
	High	0	0	0	1 (0.1)	7(1.0)	8(1.1)
	n (%)	125 (17.5)	424 (59.2)	100 (14.0)	60 (8.4)	7(1.0)	716

Table 3.

Number and percentage of women without a cancer diagnosis by risk category using self-reported and estimated weights. Cells shaded green represent women for whom the risk category remained unchanged when using estimated weight; those shaded pink indicate a single category increase and those shaded blue indicate a single category decrease using estimated weight.

		Risk using estimated weight
		Below Average	Average	Above Average	Moderate	High	n (%)
Risk using self-reported weight	Below Average	6201 (15.6)	945 (2.4)	0	0	0	7146 (18.0)
	Average	377 (0.9)	21762 (54.8)	487(1.2)	0	0	22626 (57.0)
	Above Average	0	425 (1.1)	5315(13.4)	231 (0.6)	0	5971 (15.0)
	Moderate	0	0	175 (0.4)	3231 (8.1)	58 (0.1)	3464 (8.7)
	High	0	0	0	32 (0.1)	476(1.2)	508 (1.3)
	n (%)	6578 (16.6)	23132 (58.2)	5977(15.0)	3494 (8.8)	534 (1.3)	39715

Table 2 shows that 5.2% (37 women) with a cancer diagnosis increased by a single risk category, 2.1% (15 women) dropped to a lower category. The most frequent changes (38.5%) were between below average and average risk. For the remaining 92.7% (664 women), the risk category remained unchanged. The weighted kappa statistic was 0.91, which is generally considered to show excellent agreement⁷. No women moved more than one risk category.

Change in risk categories for women who did not have a cancer diagnosis is shown in Table 3. 4.3% (1721 women) moved up a risk category while 2.5% (1009 women) decreased by a single category. The most frequent changes (34.6%) were between below average and average risk. For 93.1% of women, the risk category remained the same. The weighted kappa statistic was 0.92, suggesting there is very good agreement between risk categories. Again, no women moved more than one risk category.

Figure 2 shows the actual change in Tyrer-Cuzick risk scores for those women that moved a category; the majority of women (88.5% of those with a cancer diagnosis and 85% of those without) moved by ≤ 0.5% of their initial 10-year risk after estimating their weight from breast fat volume.

Figure 2. Actual change in Tyrer-Cuzick risk score for those women that moved a category for those with and without a cancer diagnosis.

In this cohort, neither self-reported weight nor weight estimated from the mammograms was significantly related to risk of developing breast cancer, with an odds ratio per standard deviation of 0.96 (95%CI 0.89-1.02) for self-reported weight and 1.03 (95%CI 0.96-1.11) for estimated weight.

Discussion and Conclusions

The linear regression scatter plot shows that there is a clear positive correlation between breast fat volume and self-reported weight. The correlation is also statistically significant. It is apparent from the scatterplot, that for women with self-reported weights in excess of approximately 140kg (at the upper end of the obese range) prediction of self-reported weight using breast fat volume is poor. This is also in agreement with an existing study⁵. Although that study concluded that volumetric breast density methods could not be used to predict weight, the sample size was significantly smaller (6,898 women as opposed to 40,431 women in the current study). The conclusion of the work described here differs from that study and we believe that estimating weight from mammograms provides a pragmatic solution for risk prediction, particularly where women are reluctant to be weighed or to disclose their weight. However, none of the women in PROCAS were weighed using calibrated scales, so we do not know the ground truth, and the literature suggests that self-reported weights are less reliable for heavier women⁵. Having established a method for estimating weight, it is now necessary to verify it against weights obtained using scales. The results presented in tables 1 and 2 are encouraging, showing only a small proportion of women moving risk category, and nobody moving more than one category. Figure 2 shows that of those who moved a risk category the majority changed risk by only a small amount after weights were estimated from breast fat volume. This is the case for both the women with a diagnosis of cancer and for those without. This is perhaps unsurprising given that the contribution of BMI to the Tyrer-Cuzick risk score is relatively small compared to other factors³, but such a method could prove useful where women do not provide data in relation to their weight or height. In the PROCAS study this was approximately 6.6%.

As risk prediction may be used to adjust screening intervals, of particular interest are the “moderate” and “high” categories. These women may benefit from more frequent screening than those at lower risk, and from risk-reducing interventions. The “below average” category are also of interest, for which longer screening intervals may be appropriate, or depending on the risk score, no screening at all. From the cross tabulations it is evident that the impact of using estimated weight on tailored screening would be low, since the majority of women remain in or close to the risk categories calculated using self-reported weight.

The main strength of this study lies in the size of the cohort. A disadvantage is that we were unable to determine how accurate the self-reported weights are, since weighing women attending for screening mammography is not practical on mobile screening units with a high throughput.

In conclusion, the use of estimated weights from breast fat volume looks promising but needs verification against actual weight measurements.