What question did this study set out to answer?

This study aims to evaluate a handheld bioimpedance device for distinguishing benign from cancerous breast lesions.

February 19, 2026

Abstract PS1-06-28: Non-invasive classification between benign and cancerous breast lesions via time-domain bioimpedance sensing in a first-in-human study

Key Points

This study aims to evaluate a handheld bioimpedance device for distinguishing benign from cancerous breast lesions.
Participants underwent scanning with a TD-BIS device at predefined breast locations before biopsy.
Data from 7 standard scan positions were analyzed using machine learning algorithms.
Various statistical metrics, including accuracy and predictive values, were calculated to evaluate model performance.
62 breasts were classified as cancerous or benign based on biopsy results, with 30 cancerous and 32 benign lesions.
The close scan model performed best with a balanced accuracy of 0.651 and a false negative rate of 0.039.
For combined data, the balanced accuracy dropped to 0.567, highlighting the importance of spatially relevant data.

Abstract

Abstract Introduction: Breast cancer is the most common cancer in women worldwide. While mammography remains the standard for early detection, radiation, discomfort and cost warrant development of novel screening tools and diagnostic approaches. One such approach is bioimpedance scanning, a non-invasive modality which measures electrical differences between healthy and cancerous tissue in the breast. In this study, we present a novel handheld device using time-domain bioimpedance sensing (TD-BIS) that infers dielectric properties of tissue from voltage time-series data. Analysing this data with machine learning methods, the TD-BIS technology was used to determine if lesions present in patients were either benign or cancerous in nature. Methods: 68 participants undergoing routine mammography or breast ultrasound were enrolled post radiological review and diagnosed by biopsy. Prior to the biopsy, participants were scanned on each breast using the TD-BIS device at predefined locations on the breast. Scans from 7 standard positions were recorded from each breast, starting with one taken directly on the nipple (N) and the remaining scans taken at clock-face positions 2, 4, 6, 8, 10, and 12, with the nipple at the centre. For the right breast, these positions were labelled ‘inner above (IA), 'inner below’ (IB), ‘below’ (B), ‘outer below’ (OB), ‘outer above’ (OA), and ‘above’ (A) respectively. OA/IA and OB/OB were swapped for the left breast. The range, mean, standard deviation, skewness and kurtosis for different groups of capacitances were used as input for a random forest model. Distinguishing between cancerous and benign lesion classification, the F1 score, balanced accuracy (BA), false positive rate (FPR), false negative rate (FNR), positive predictive value (PPV) and negative predictive value (NPV) were calculated. Models were trained using either one scan per patient from the above list (single scan), all scans combined (combined scans), or all scan data that was within 8 cm of the known lesion (close scans). Mann-Whitney tests (with a Bonferroni correction) were used to evaluate the significance between models. Independent t-tests were used to compare the performance of each scan position relative a random permutation of the labels. Results: 62 breasts were categorized as cancerous (n=30) or benign (n=32) via radiological and biopsy evaluation. Results showed that the most significant and best (according to BA and F1) performing models in cancer/benign distinction were ‘Close’, A, IA, IB, N, and B models. However, using only one scan position resulted in high FNR. For close scan data, 12 of 30 benign lesions and 31 of 32 cancerous lesions were correctly identified, respectively. The performance for close scans was: BA:0.651, F1:0.623, FNR:0.039, FPR:0.660, PPV:0.609, NPV:0.895, all with p-value 0.05. when compared to Combined. When evaluating the combined data, 6/30 benign lesions and 31/32 cancerous lesions were correctly identified, with BA:0.567, F1:0.497, FNR:0.031, FPR:0.834, PPV:0.553, and NPV:0.828. Discussion: This initial evaluation of TD-BIS demonstrates its feasibility as a method for discriminating between benign and cancerous lesions. When evaluating spatially relevant data, the study shows that data closest to the lesion improves the cancer detection models when compared to non-spatial data. This implies that the TD-BIS device measures biologically relevant dielectric properties of the underlying tissue. The performance of the close model reveals an FNR less than 4%, suggesting this technology has the potential to be a non-invasive clinical tool to discriminate between benign and cancerous lesions. This conclusion is confirmed when the model input is expanded to include scans farther from the lesion site (e.g combined), which significantly reduces model performance. Citation Format: M. Goldfinger, J. Fernandez Vargas, L. Honeyfield, C. Colcol, K. Wourms, M. Taani, S. Flais, A. Antrobus, A. Lim. Non-invasive classification between benign and cancerous breast lesions via time-domain bioimpedance sensing in a first-in-human study abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS1-06-28.

AI से पूछें

Bookmark