Breast histoscanning: the development of a novel technique to improve tissue characterization during breast ultrasound
- 890 Downloads
KeywordsMalignant Lesion Papilloma Training Dataset Normal Volunteer Preoperative Assessment
Imaging alone cannot reliably distinguish benign/malignant breast disease or assess the extent of cancer. This study assesses the feasibility of using additional information obtained at US (BHS) to aid diagnosis and preoperative assessment.
3D US scans at 8 MHz, 12 MHz, 15 MHz were obtained of breast tissue in normal volunteers in two planes and with/without harmonics. Five volumes of sagittal scans at 8 MHz from three individuals were used to identify normal characteristics and define the baseline. The 3D volume was divided into voxels (0.1 x 2 x 1.5 mm) and raw data from each voxel were analysed by applying linear and nonlinear classifiers to assess 29 statistical characteristics (BHS). The training dataset contained 300,000 voxels. After training, the classifier’s output showed 3% error on both normal and abnormal tissue. The algorithm was tested on 32 further volumes representing 6,000,000 voxels of normal and abnormal tissue from 20 individuals. Abnormal tissue included various biopsy-proven lesions: malignancy (six), papilloma (one), hamartoma (one), fibroadenoma (two), cyst (two), fibrosis (one). Subclassifiers were developed to distinguish between cancer and benign voxels.
In 17 normal testing volumes, 3% of isolated voxels were classified as abnormal. In 15 abnormal testing volumes, the subclassifiers differentiated between malignant and benign tissue. BHS in benign tissue showed <1% abnormal voxels in cyst, hamartoma, papilloma and benign fibrosis. The fibroadenomas differed showing <5% and <24% abnormal voxels. Abnormal voxels in cancers increased with the volume of cancer at pathology.
Histoscanning reliably discriminated normal from abnormal tissue and could distinguish between benign and malignant lesions.
This article is published under license to BioMed Central Ltd.