Preoperative MRI has been shown to have a crucial role in the assessment of breast cancer patients potentially eligible for NSM. Compared to mammography, which is a 2D imaging modality, breast MRI provides a 3D evaluation of the whole breast, thus reducing the loss of spatial information about tumour extent and location. For that reason, MRI has been considered as the method of choice to preoperatively predict occult nipple involvement [20, 21].
To facilitate surgical planning and to standardise the use of tumour-to-NAC distance as the main predictor of nipple infiltration by tumour, Giannini et al. [16] recently developed an automated method to compute the 3D tumour-to-NAC distance, which overcomes the performance of manual 2D methods in predicting NAC involvement. However, this method was developed and validated using images acquired with the same MRI scanner and having the same acquisition protocol. When developing automated methods, this could represent a strong bias, since images strongly differ between scanners and imaging protocols.
In the current study, we validated this algorithm using an external dataset of images acquired in a different centre, using a different MRI scanner and acquisition protocol from that used in the development phase. The performance reached by this method with this external dataset (sensitivity 78%, specificity 72%) demonstrated that this 3D automated method could represent a reliable method to preoperatively compute tumour-to-NAC distance, improving the management of patients candidate for NSM. The algorithm had a failure rate of only 5% because of the failure of the nipple (two cases) or tumour (two cases) segmentation.
The automated system presented in this study may show many advantages. First, the 3D tumour-to-NAC measurements were more reliable than the 2D measurements calculated using MIP images. In fact, when a measurement is carried out on the axial/sagittal projection, the information along the z-/y-axes is lost, with the consequent chance that the lesion and the nipple appear closer, as lying on the same x-/y-axis. Actually, the lesion and nipple are often more distant, as they are seated in different slices of MRI volume. In fact, the distance calculated by the automated algorithm was greater than that manually calculated in 55/72 cases (76%). In a recent study [21], the issue of three-dimensional “real” distance has been discussed. However, in that case, the measurements were done in a completely manual way, by computing four distinct distances in each case, using digital images on flat-screen liquid-crystal display monitors. This is a time-consuming task, which is difficult to apply in clinical practice. In addition, this study [21] did not make a comparison with the standard methods.
Interestingly, no fully automated methods for the nipple segmentation on MRI are available yet. The nipples often differ in form and intensity of the signal in different patients; in addition, the nipple is not always perfectly located at the centre of the T2-weighted image and, when inverted, cannot always be distinguished from glandular tissue. In our experience, 92% of the nipples were properly segmented by the algorithm.
Taken for granted that the tumour-to-NAC distance is up to date the most useful parameter for the preoperative assessment when NSM is under consideration, the main issue is to define the best cut-off value capable of predicting NAC involvement. In this regard, the literature is inhomogeneous. Some authors propose 10 mm as the ideal cut-off [13], while others recommend 20 mm [22, 23]. In a recent study, a distance of 5 mm was suggested [14].
In the present series, the 3D automated method improved the diagnostic performance when compared to 2D manual measurement, even though not significantly. In particular, the best compromise between sensitivity and specificity for each method was reached using the cut-off of 30 mm for the automated method and of 21 mm for the manual method. This difference is consistent with the increase by 11.5 mm in the average tumour-to-NAC distance when processed by the automated method versus the manual one and with the previously mentioned greater distance in 76% of cases as compared to the manual measurement.
As shown in Table 2, specificity and PPV of the automated method overcome those of the manual method at all the cut-off values. Sensitivity and NPV are instead higher only for the best cut-off (30 mm) since the automated method is not able to clearly identify NAC-negative patients at smaller distances. All the patients with tumour-to-NAC automated distance ≤ 5 mm and ≤ 10 mm showed tumour involvement of the nipple at the final pathology, confirming the high specificity of the automated method. This performance is higher compared to the manual measurement, which showed 96% and 86% specificity at these thresholds. However, the sensitivity at ≤ 5 and ≤ 10 mm was very low: only 9% and 26% of the patients with tumour involvement of the nipple were positive when tested with the automated method at these cut-off values, respectively.
Since the aim of the assessment before surgery is to propose NSM to all patients who may potentially preserve the nipple (i.e. patients without NAC tumour involvement at pathology), specificity and PPV are the most useful preliminary parameters to know. As a high specificity is related to a low sensitivity, many patients with NAC involvement (i.e. patients with NAC tumour involvement at pathology) will still be candidates for NSM.
By choosing the cut-off at ≤ 10 mm, the high specificity (100%) allows to exclude all the patients who certainly will not be able to keep the nipple, while the low sensitivity (26%) causes the inclusion in the selection for NSM of most patients (74%) with NAC involvement. However, the current protocol for the NSM mandates the intraoperative histological examination of the retroareolar tissue, which shows a good negative predictive ability. In such cases, the surgery may be converted to SSM at the same surgical time.
The data obtained from our study are similar to those we have previously obtained with the same 3D methods on images obtained on a different equipment, as shown in Table 3, as expected from an automated not operator-sensitive and therefore more reliable algorithm. The main difference is between the value of the 3D best cut-off considered in the previous and in the present study, 21 mm and 30 mm, respectively. This is an important issue that deserves further investigation. A prospective series taking into account the breast volume as well as the use of a standardised MRI protocol for the acquisition of T2-weighted and DCE images may help overcoming the gap.
Limitations of our study are mainly related to its retrospective design. Our series covers a period of almost 6 years, during which there have been technological advances that, although modest, could have influenced the signal to noise ratio and the image quality, changing the segmentation capabilities (especially for the nipples) of the automated algorithm, as shown in Fig. 6. The incomplete segmentation of the NAC causes the automated distance to be computed between the edge of the lesion and the front part of the nipple, rather than between the edge of the lesion and the base of NAC. Moreover, the correct localisation of the nipple could be difficult in some cases, such as malposition or introflexion, and need the supervision of a technician with specialised experience in breast MRI. The implementation of this step is definitely needed to maximise the accuracy of the algorithm. Secondly, choosing the cut-off at ≤ 10 mm, the sensitivity of the automated method remains low and many positive cases will be overlooked. It is, therefore, necessary to perform the frozen examination of subareolar tissue during surgery, which is currently considered the safest method to predict nipple involvement.
In conclusion, our study suggests that breast MRI is a promising method for the preoperative assessment of patients candidate for NSM by predicting the occult involvement of the NAC. Our novel 3D automated method seems to improve the results obtained with the 2D manual distance measurement also being validated on an independent external dataset. A cut-off value of ≤ 10 mm provided great accuracy as all the patients with a tumour-to-NAC distance ≤ 10 mm require the removal of the nipple. If integrated into clinical practice, this method could be useful to reduce the variability in selecting patients who may have the nipple preserved.