For in-vitro test, the mismatch on the reference amorphous lesion identified in prone MRI and the one visualized in real-time US, calculated between the real-time US probe position and the MRI set target, was minimal (Fig.6). No further investigations or statistical analysis has been performed, since the phantom deformation is limited, compared to real tissue, therefore phantoms are not an adequate representative of the real breast organ deformation.
For in-vivo tests, 2 different investigation scenarios were performed.
The first investigation was made to evaluate the BreastNav® feasibility and its AI algorithm performance on real patients. In a previous study, the algorithm has been already validated by means of MRI breast examinations, acquired in prone and supine positions.
One healthy volunteer (age:40y) with no lesions and hypertrophic breast has been involved.
Before MRI acquisition, 8 Skin Landmarks (SL) have been positioned on each breast, 5 used for registration purposes (corresponding to P1-P5 points) and 3 as target SL, placed in the position at 3, 6 and 9 o’clock in sub-areola position. These SL better describe the organ deformation and slipping, due to prone-supine change and they are used to compute the error between their position during supine US and the one computed by the AI algorithm, based on MRI acquisitions.
MRI acquisition (1.5T Ingenia-Philips) was performed in prone (4-mm axial T2W-TSE with and without fat suppression,DWI b-values 0-500-1000s/mm2,DCE-Gadobutrol;0,05 mmol/kg;2ml/s) and supine position (4-mm axial T2-SSH with and without fat suppression MultiVaneXD HR) and only the prone DICOM sequence loaded on US system and registered in BreastNav environment; the Euclidean error between the real-time US probe position and the MRI set target, are shown in Table 1.
The second investigation involved 4 patients (medium age: 55y) with MRI reference lesions and no SL positioned.
These patients came from the routine MRI, due to different clinical situations/settings, where a second-look US has been prescribed.
MRI acquisition (GE-1.5T) was performed only in prone position (Axial-T1,STIR,DWI,DCE 1+5-ProHance) and DICOM data loaded on US system; in BreastNav environment the lesions in MRI BI have been identified and target.
The following clinical cases were analyzed: highly suspicious wide area of micro-calcification on external quadrants of left breast; neoplastic mass on upper quadrants of right breast; US follow-up of architectural distortion with correlated MRI contrast enhancement; the last, breast prosthesis and parenchymal mastopathy area on left breast with correlated MRI enhancement.
The lesions in the 4 patients analyzed have a consistent dimension, superior to 1 cm: this choice has been made in order to easily identify the reference target during US and to properly test the feasibility of the technology on real patients.
In all patients, Breast Nav technology has been able to make prone MRI-US supine FI, identifying properly the lesion during US, previously marked in MRI, with a mismatch between a minimum of 4 mm to a maximum of 18 mm, corresponding to the prosthesis case, caused by a different behavior compared to the real organ during patient’s position changing (Fig.7-8).
Benefits
Breast Nav technology allows to save all data in the US archive for post-processing analysis and review, to print pictures of the target reference and US probe spatial position with FI and to print reports, also including BI-RADS® categorization (Breast Imaging Reporting and Data System-ATLAS of the American College of Radiology).
During BreastNav MRI-US FI investigation other breast relevant US technologies such as microvessel imaging, microenhancement imaging and elastography (Strain and/or Shear Wave technology) can be also used, making the method a real multimodality BI approach (Fig. 9-11).
Limitations
The algorithm performance provided good results in this preliminary study, but the patients sample is limited; breast size, its tissue composition and the lesion position can affect the final accuracy, due the fact that the deformation and the mammary gland slipping are not homogeneous among different patients and the variability observed among women is huge.
For this reason, further studies are necessary to optimize the system accuracy, improving the machine learning procedure by increasing the big data cases, in order to provide a more diagnostic confidence.