Diffuse reflectance spectroscopy for enhanced diagnostic precision in breast cancer Authors: Lianne Feenstra, Marcos Da Silva Guimaraes, Caroline A. Drukker, Frederieke van Duijnhoven, Behdad Dashtbozorg, Theo J. M. Ruers Published: 2025 ABSTRACT PurposeVacuum-assisted breast biopsy (VABB), widely used for diagnosing breast cancer, is prone to sampling errors that may necessitate repeat procedures and delay treatment. Optically guided VABB using Diffuse Reflectance Spectroscopy (DRS) may help address this limitation.MethodsTo this end, an optical introducer (a cut-out 9-gauge hollow tube) was developed and equipped with optical fibers. The tube slides over a standard VABB needle, enabling tis- sue measurements near the biopsy aperture. Ex vivo DRS measurements were performed on lumpectomy specimens from 90 patients using this tube, yielding over 500 measurement loca- tions. Tissue labels were derived through deformable image registration of measurement sites with annotated histology, and label extraction was optimized for larger fiber distances. Machine learning classifiers were trained on the spectral data.ResultsEnsemble models achieved up to 94% sensitivity at a 20% tumor threshold and maintained high performance at lower thresholds.ConclusionThese findings demonstrate that DRS enables real-time tissue characterization at large fiber distances and may improve biopsy precision. In vivo studies are needed to validate performance and support the integration of DRS into routine VABB diagnostic workflows. Introduction Vacuum-assisted breast biopsy (VABB) is a minimally invasive technique that has become a key component in the diagnostic workup of suspicious breast lesions, offering improved accuracy and reliability compared to core needle biopsy (CNB) [1, 2]. By using a vacuum mechanism to collect larger, contiguous tissue samples, VABB minimizes sampling error and reduces the underestima- tion of conditions such as atypical hyperplasia, ductal carcinoma in situ (DCIS), and small invasive cancers. Its compatibility with multiple imaging modalities, including ultrasound, stereotactic mam- mography, and MRI, makes it highly versatile and especially valuable for lesions that are not visible on ultrasound or are located in dense breast tissue [3]. VABB is particularly effective for sampling microcalcifications, architectural distortions, and other non-palpable abnormalities, often enabling more definitive diagnoses in these cases and reducing the need for surgical excision [4]. Compared to CNB, VABB provides superior tissue yield, improved diagnostic accuracy, and broader applicability across lesion types and imaging settings, making it the preferred method in many clinical scenarios. Obtaining representative tissue samples is essential for accurate breast cancer diagnosis, yet cur- rent VABB procedures still face notable limitations. VABB, which typically yields six to nine tissue samples [5], provides larger specimens than CNB but still carries a risk of sampling errors. Under- sampling, when the most representative part of the lesion is missed, can lead to incomplete or non-diagnostic samples contributing to underestimation of malignancy and delayed or incorrect diag- noses [2, 6, 7]. To compensate, radiologists may extract more tissue than necessary, which improves diagnostic confidence but raises concerns about oversampling, patient discomfort and risk of com- plication such as bleeding. These challenges highlight the critical importance of accurate tissue sampling, which relies heavily on precise needle placement. However, standard imaging techniques such as ultrasonography (US) and stereotactic mammography offer only 2D views of a 3D anatomy [8, 9], increasing the risk of missing key areas within the suspicious area. Even minimal deviations in angle or depth can compromise sample quality [10, 11]. Consequently, VABB may need to be repeated, with re-biopsy rates reaching up to 9%, particularly in cases involving lesion heterogeneity or radiologic-pathologic mismatch [12, 13]. These repeat procedures delay diagnosis, increase patient anxiety, and add to healthcare costs. Improving the accuracy of targeting and sampling strategies is therefore critical to enhance diagnostic reliability and reduce the need for repeat procedures. One way to improve the accuracy of VABB is by integrating real-time tissue assessment into the procedure. Optical techniques, such as diffuse reflectance spectroscopy (DRS), offer a practical solu- tion for this purpose [14]. DRS measures reflected light from tissues to evaluate the tissue properties, offering immediate information about the biochemical composition [15]. It has been demonstrated that DRS shows a high sensitivity in distinguishing between benign and malignant breast tissue [16–19]. To allow more representative biopsies, optical fibers can be attached to a biopsy needle to assess tissue characteristics before sampling. A positive signal could indicate that the needle is within a tumor, while a negative reading may suggest the need to reposition the needle along its path. This real-time feedback enables clinicians to target suspicious areas better and avoid non-representative samples. By incorporating DRS into the biopsy workflow, clinicians can make informed decisions on needle placement during the procedure, potentially reducing repeat biopsies and improving diagnos- tic accuracy. Earlier DRS-based biopsy approaches have primarily used fiber optics positioned at the needle tip, providing diagnostic information but sampling tissue volumes that do not fully overlap with the biopsy site itself [20, 21]. This study addresses the challenges of incorporating DRS into the VABB procedure. We intro- duce a different design concept to allow more relevant biopsies: an optical introducer (a cut-out hollow tube equipped with optical fibers) that slides over a standard VABB needle, enabling DRS measurements directly along the biopsy aperture. This design ensures that the optical assessment corresponds to the same tissue volume that will be excised, improving diagnostic reliability. Since most earlier studies focused on smaller needles with short fiber distances [20, 21], it remains unclear whether accurate tissue classification can still be achieved under these new conditions. Defining the appropriate measurement volume for this increased fiber distance is critical for accurate tissue labeling and developing robust tissue classification algorithms. We aim to optimize the label extrac- tion by considering the measurement volume based on fiber distance using our existing pipeline for registering and labeling optical measurements [22, 23]. Another concern is the position of the optical fibers along the biopsy device. Placing the fibers close to the biopsy aperture may improve targeting precision but risks signal distortion caused by nearby metallic components, potentially lowering classification accuracy. While optical fibers posi- tioned at the tip of a needle have shown high diagnostic accuracy [24], it remains uncertain how fiber placement along the biopsy needle, especially near the biopsy aperture, influences optical mea- surement performance. To address this, phantom measurements designed to assess the influence of fiber positioning and metal interference on signal quality will be conducted. Our ultimate goal is to assess whether DRS, even with increased fiber distances and near the biopsy aperture where metallic components may influence measurements, can still accurately distinguish healthy tissue from tumor tissue, including invasive carcinoma (IC) and DSIC. If feasible, this approach could enable more targeted sampling of potentially cancerous tissue while minimizing the removal of healthy tissue, thereby enhancing biopsy accuracy in breast cancer diagnosis. Materials and methods Materials Diffuse reflectance spectroscopy device The DRS device consists of an optical introducer and a measurement console. The DRS console con- sists of a halogen broadband light source (Avantes AVALIGHT-HAL-SMINI) and two spectrometers: one for the visible range (200–1160 nm, AVASPEC-HS2048XL-EVO) and one for the near-infrared range (900–1750 nm, AVASPECNIR256-1.7-RS). Custom MATLAB software controls the system and merges the spectra into a single 400–1600 nm range. The optical introducer (a cut-out 9-gauge hollow tube) includes two 400 μm core fibers with 4.7 mm fiber distance. One fiber delivers light from the halogen source, the other collects reflected light. The optical introducer fits over a 9-gauge VABB needle, enabling DRS measurements around the needle aperture (Fig. 1). Before measure- ments, spectra were calibrated using a white reference (Spectralon, Avantes WS-2) to compensate for variations in lamp intensity and the system’s wavelength-dependent response. A dark reference (with the lamp turned off) was also recorded to account for detector dark current and to reduce artifacts from ambient light. The final calibrated reflectance spectrum was calculated using:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R_{{\rm{cal}}}} = {{{R_{{\rm{meas}}}} - {R_{{\rm{dark}}}}} \over {{R_{{\rm{white}}}} - {R_{{\rm{dark}}}}}}$$\end{document} where Rcal is the calibrated measurement, Rmeas is the raw measurement, Rdark is the dark reference, and Rwhite is the white reference. Fig. 1Overview DRS device: the optical introducer containing two optical fibers (4.7 mm fiber distance). An illuminating fiber is connected to a light source, while the detection fiber is connected to spectrometers. Custom software is used to display the reflectance spectrum. The optical introducer (b) can be placed over an original VABB needle (a), enabling optical measurements along the needle aperture (c) Overview DRS device: the optical introducer containing two optical fibers (4.7 mm fiber distance). An illuminating fiber is connected to a light source, while the detection fiber is connected to spectrometers. Custom software is used to display the reflectance spectrum. The optical introducer (b) can be placed over an original VABB needle (a), enabling optical measurements along the needle aperture (c) Method This study investigates the use of optical fibers along a VABB needle aperture to enable tissue char- acterization of IC and/or DCIS using DRS. First, we will assess the design of the optical introducer, including the VABB needle, by evaluating signal distortion introduced by metallic components of the needle. Next, we describe a pipeline to acquire and label optical data from ex vivo breast tissue for training a classification algorithm. Special attention is given to optimizing label extraction for larger fiber distances. Metal interference assessment Positioning optical fibers close to the biopsy needle’s aperture can support more precise tissue tar- geting and sampling, but presents challenges due to potential interference from metal components. Light interacting with the metal can cause signal distortion, attenuation, and noise, compromis- ing tissue classification accuracy, especially since classification algorithms are typically developed without accounting for such interference. We evaluated the impact of metal components on optical signals using a standardized phantom, prepared with a 1/30 dilution of intralipid and water to simulate tissue properties. The introducer (Fig. 1b) was secured in a fixed holder and placed in the intralipid to perform DRS measurements. Subsequently, the introducer was positioned over a 9-gauge VABB needle (Fig. 1c), also held in place within the intralipid. Several DRS measure- ments were taken at various positions along the aperture. We applied Standard Normal Variate (SNV) normalization to reduce multiplicative effects from metal-induced reflectance, enhancing signal consistency, and improving classification reliability [25]. For DRS, the SNV transformation is computed using the following equation:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{SNV(}}{\lambda _i}{\rm{)}} = {{{\rm{R(}}{\lambda _i}{\rm{)}} - {\rm{\bar R}}} \over {{\rm{std}}\left({\rm{R}} \right)}}$$\end{document} where: R(λi) is the reflectance value at wavelength λi,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{\bar R}}$$\end{document} is the mean reflectance value across all wavelengths,std(R) is the standard deviation of the reflectance values across all wavelengths. R(λi) is the reflectance value at wavelength λi, R(λi) is the reflectance value at wavelength λi, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{\bar R}}$$\end{document} is the mean reflectance value across all wavelengths, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{\bar R}}$$\end{document} is the mean reflectance value across all wavelengths, std(R) is the standard deviation of the reflectance values across all wavelengths. std(R) is the standard deviation of the reflectance values across all wavelengths. Ex vivo lumpectomy study We performed DRS measurements on freshly excised breast lumpectomy specimens using the optical introducer to investigate the tissue discrimination capability for a larger fiber distance. The mea- surement and labeling process has been detailed in previous studies [22, 23]. A summary is provided here. Patient inclusion The study included 90 breast cancer patients scheduled for breast-conserving surgery at the Antoni van Leeuwenhoek Hospital between June 2021 and September 2023. Eligible patients had a diagnosis of IC and/or DCIS based on preoperative examinations. Patients were included regardless of neo- adjuvant treatment. The study was approved by the Ethics Review Board (IRBm 20 − 006), and all patients gave informed consent for the use of their data and tissue for research. DRS measurements After breast-conserving surgery, the excised tissue (lumpectomy specimen) was collected, inked, and sectioned until tumor tissue became visible. The unsliced half of the specimen was mounted in a holder and placed under the Point Projection Mapping (PPM) system, where points of interest (POIs) were manually selected, projected onto the specimen, and captured in images. On each POI separately, three DRS measurements were performed and averaged. After measurements, the specimen was fur- ther processed into hematoxylin and eosin (H&E) sections, digitized, and annotated by a pathologist. Snapshot images were collected to correlate the optical measurements with histopathology, ensuring accurate labeling of the optical data [22] (Fig. 2). Fig. 2Optical measurement pipeline and correlation with pathology. a. The PPM system projects measurement locations onto the surface of the half-sliced lumpectomy specimen, and DRS measure- ments are performed at each point. b. The specimen is processed in H&E sections, and a pathologist annotates the measured slice. c. Deformable image registration is used to register the specimen image with its corresponding histology image Optical measurement pipeline and correlation with pathology. a. The PPM system projects measurement locations onto the surface of the half-sliced lumpectomy specimen, and DRS measure- ments are performed at each point. b. The specimen is processed in H&E sections, and a pathologist annotates the measured slice. c. Deformable image registration is used to register the specimen image with its corresponding histology image Correlation with histopathology A previously developed, deep-learning-based deformable multi-modal image registration method was used to align the specimen image with its corresponding histology image, accounting for tissue defor- mations between images from different modalities [23]. The model produces a dense displacement field (DDF) that maps coordinates from one image to another image by defining a displacement vector for every point. Applying the DDF transformation on the measurement location image enables accu- rate mapping of optical measurements onto the ground truth histology image and label extraction as described in the following section. Label extraction In this study, a large fiber distance between the light source and detector enables deeper light pen- etration and a broader spread within the tissue, capturing signals from a larger tissue volume that extends beyond the physical size of the optical probe. Therefore, determining the correct measure- ment volume is essential for assigning a label to each spectrum (in our case, the histology slice from the measured lumpectomy surface), as it reflects the tissue percentage labels. For label extraction, measurement locations from the specimen image with POIs were overlaid with the annotated histol- ogy image (as ground truth). Therefore, coordinates of the measurement locations were identified, and ellipsoids (representing the diffuse measurement volume) were generated. To bring these loca- tions to the correct orientation, the DDF transformation is applied to these extracted ellipsoids. Afterwards, the registered ellipsoids were overlaid on the annotated histology image, making it pos- sible to calculate tissue-type percentages at each measurement location [22]. To evaluate labeling accuracy for tissue classification using larger fiber distances, six separate datasets were created each using a different ellipsoid size with diameters ranging from 4 to 9 mm (Fig. 3). Fig. 3Overview label extraction: a. Specimen image with POIs. b. Binary image with extracted POIs represented as ellipsoids (ranging from 4–9 mm). c. The DDF acquired from deformable image registration is applied to the ellipsoids in (b), resulting in registered POIs (ranging from 4–9 mm). d. Histology image with tumor tissue annotated by a pathologist, serving as ground truth. The histology image is thresholded, assigning colors to each pixel: red for IC and DCIS, green for connective tissue, and yellow for fat tissue. e. The registered POIs (c) are overlaid with the thresholded histology image (d), creating six distinct ellipsoid datasets, each with varying tissue percentages per measurement location used for further classification Overview label extraction: a. Specimen image with POIs. b. Binary image with extracted POIs represented as ellipsoids (ranging from 4–9 mm). c. The DDF acquired from deformable image registration is applied to the ellipsoids in (b), resulting in registered POIs (ranging from 4–9 mm). d. Histology image with tumor tissue annotated by a pathologist, serving as ground truth. The histology image is thresholded, assigning colors to each pixel: red for IC and DCIS, green for connective tissue, and yellow for fat tissue. e. The registered POIs (c) are overlaid with the thresholded histology image (d), creating six distinct ellipsoid datasets, each with varying tissue percentages per measurement location used for further classification Data processing Here, we describe a method to determine the optimal label size for label certainty and subsequent breast tissue classification based on spectral data acquired with a large fiber distance. Four classifiers were selected to capture varying levels of model complexity and interpretability. Support vector machine (SVM) and k-nearest neighbors (KNN) were included as established baseline models for spectral data classification, providing interpretable decision boundaries and reliable performance on smaller datasets. Ensemble (ENS) and neural network (NN) classifiers were chosen to evaluate whether non-linear, data-driven algorithms could better capture complex spectral differences between healthy and tumor tissue. This combination enabled a balanced comparison between conventional and advanced learning approaches for breast tissue spectroscopy, as commonly explored in DRS- based analyses [26, 27]. Bayesian optimization was applied uniformly across all four classifiers to ensure a consistent and unbiased model comparison. The same optimization framework and objective function (cross-validated classification accuracy) were used for all models, with parameter search ranges tailored to the relevant hyperparameters of each algorithm. All analyses were performed in MATLAB R2024a (MathWorks, Natick, MA, USA). An overview of the data analysis steps can be found in Fig. 4. Spectra preprocessing and normalization Before classification, wavelengths below 500 nm and above 1600 nm were excluded due to low signal- to-noise ratios. To correct for intensity variations caused by scattering, path length differences, and potential metal interference, spectra were normalized using SNV (Eq. 2). This preprocessing step also helped reduce inter-patient variability and background noise. Fig. 4Overview data analysis steps: The acquired data was preprocessed first. The normalized and scaled spectra were used for further tissue classification. This involves optimizing label extraction for larger fiber distances. Next, different models were trained and evaluated Overview data analysis steps: The acquired data was preprocessed first. The normalized and scaled spectra were used for further tissue classification. This involves optimizing label extraction for larger fiber distances. Next, different models were trained and evaluated Optimal label extraction To evaluate the certainty of extracted labels, six datasets were created using ellipsoids with diam- eters ranging from 4 to 9 mm. Each dataset contained ground truth tissue percentages for IC, DCIS, connective tissue, and fat as described in Section "Ex vivo lumpectomy study". The goal was to identify the ellipsoid size that best represents the measurement volume for larger fiber distances, enabling reliable labeling of the acquired DRS data. Therefore, label consistency for each measurement location was assessed by calculating the total tumor percentage (IC + DCIS) across all ellipsoids for all six different sizes. The standard deviation (SD) of these percentages was used to assess how much the tumor tissue varied within each measure- ment location. A low SD reflected uniform tissue composition (regardless of ellipsoid size) and was used to select measurement locations with the most certain labels for the training set. In contrast, measurements with higher variation, defined as an SD of ≥ 0.05, were included in the test set to challenge the classifier under more diverse labeling conditions. Each ellipsoid dataset was analyzed separately using the same locations assigned to the training and test sets. It is important to note that while the measurement locations assigned to the training and test sets remained the same across all six experiments, variations in tissue composition within the test sets could result in different assigned labels (tumor/healthy). Locations containing any tumor (tumor percentage > 0%) were assigned as tumor for the evaluation of classification performance, to avoid bias in determining the optimal ellipsoid size. Four optimized supervised classifiers: SVM, KNN, ENS, and NN, were evaluated. For each model, sensitivity, specificity, and area under the curve (AUC) were calculated and averaged per dataset. The dataset yielding the highest overall performance was selected for further classification analysis. This approach may lead to lower performance metrics, which is addressed in Section "Data Processing" where we evaluate classification performance based on varying tumor percentages. Tissue classification To further refine tumor classification, datasets were created using varying tumor thresholds, labeling measurements with ≥ 5%, 10%, 15 or 20% IC and/or DCIS as tumor tissue. Samples below each threshold were labeled as healthy (fat and/or connective tissue). Each dataset was balanced to ensure an equal number of tumor and healthy spectra. Evaluating multiple thresholds enabled the assessment of DRS capability in detecting tumor tissue within a larger measurement volume. The SVM, KNN, ENS, and NN classifiers were trained and evaluated for each tumor thresh- old using repeated 10-fold cross-validation (10 iterations) combined with Bayesian hyperparameter optimization. This approach minimizes overfitting by generating multiple train–test sets, providing a more reliable estimate of model performance on unseen data. To further prevent overfitting in the NN, the model architecture was constrained to a maximum of three layers with up to ten neurons per layer, combined with a regularization term. Similarly, for the ENS model, the number of learners was limited to ten, and for decision tree–based learners, the maximum number of splits was restricted to thirty. Performance was assessed based on accuracy (ACC), sensitivity, specificity, and AUC, aver- aged across thresholds and iterations. The best-performing classifier was further evaluated using the Matthews correlation coefficient (MCC), which offers a balanced measure of overall classification performance to identify the optimal tumor threshold for tissue classification. Next, receiver operat- ing characteristic (ROC curves) were generated for each tumor threshold using the best-performing model. The AUC was calculated, and the optimal operating points, which minimized the average misclassification cost, were used to determine accuracy (ACC), sensitivity, and specificity. Results Metal interference assessment To evaluate the impact of metal on DRS spectra, 40 optical measurements were acquired using the optical introducer along the aperture of a VABB needle embedded in an intralipid phantom. These measurements were compared to a metal-free reference, obtained with the introducer placed in the intralipid phantom without the needle. Clear differences were observed between the unnormalized spectra from metal-affected and the metal-free reference measurements (Fig. 5a). The reference spectrum shows significantly higher reflectance values, particularly between 400 and 1200 nm. The intensity of the metal-affected measurement varies, as shown by the standard deviation. The SNV normalization improved consistency among metal-affected spectra, reducing the standard deviation (Fig. 5b). Fig. 5A reference measurement with optical introducer only (no metal) is compared with 40 measurements with the optical introducer along a VABB needle aperture (metal-affected), both performed in an intralipid phantom. The reference measurement is visualized in blue, whereas the average metal-affected measurement (and standard deviation SD) is visualized in orange. Original spectra (a) are compared with SNV-normalized spectra (b) A reference measurement with optical introducer only (no metal) is compared with 40 measurements with the optical introducer along a VABB needle aperture (metal-affected), both performed in an intralipid phantom. The reference measurement is visualized in blue, whereas the average metal-affected measurement (and standard deviation SD) is visualized in orange. Original spectra (a) are compared with SNV-normalized spectra (b) Ex vivo lumpectomy study: tissue classification Optimal label extraction The performance of four supervised machine learning classifiers: SVM, KNN, ENS, and NN was averaged and evaluated for different ellipsoid label sizes (4–9 mm). Figure 6 shows the effect of varying ellipsoid sizes on sensitivity, specificity, and AUC. As the ellipsoid size increased, both specificity and AUC improved. Specificity increased from an average of 66% at 4 mm to over 78% at 9 mm, while. AUC increased from an average 0.55 to 0.68. Sensitivity remained relatively stable across all sizes, showing only minor variation. Overall, the 9 mm ellipsoid size achieved the highest specificity and AUC values without a reduction in sensitivity, suggesting it provides the most effective balance for tumor detection in this context and will therefore be used for further analysis in this study. Fig. 6Effect of ellipsoid size (4–9 mm) on detection performance across three metrics: (a) Sensitivity, (b) Specificity, and (c) AUC. A boxplot represents the distribution of values for a given ellipsoid size, displayed in a color of blue. Boxes show the interquartile range, the horizontal line indicates the median, and whiskers represent the full data range Effect of ellipsoid size (4–9 mm) on detection performance across three metrics: (a) Sensitivity, (b) Specificity, and (c) AUC. A boxplot represents the distribution of values for a given ellipsoid size, displayed in a color of blue. Boxes show the interquartile range, the horizontal line indicates the median, and whiskers represent the full data range Data distribution Ninety patients were included in the study (Table 1), and 533 DRS measurement locations were collected. Of these, 250 (46.9%) contained tumor tissue (IC and/or DCIS), and 283 (53.1%) contained healthy tissue (connective tissue and/or fat) using the 9 mm elipsoid dataset. Figure 7 shows the tissue type distribution. We investigated different tumor thresholds (5%, 10%, 15%, and 20%) for tumor label assignment and explored their effect on the classification performance. This resulted in four separate datasets with 172, 160, 151, and 141 tumor measurements, respectively. Each dataset was randomly balanced with an equal number of healthy measurements. Fig. 7Tissue percentages of invasive carcinoma, Ductal carcinoma in situ, connective - and fat tissue for each separate measurement location, for the 9 mm elipsoid dataset Tissue percentages of invasive carcinoma, Ductal carcinoma in situ, connective - and fat tissue for each separate measurement location, for the 9 mm elipsoid dataset Table 1Characteristics of patient cohort and measurement locationsEx vivo lumpectomy specimens (N = 90)Age (years) (median, SD)56,6 (11,7)Lesion diameter (mm) (median, SD)22,6 (16,1) Neoadjuvant treatment Chemotherapy26 (28,9%) Endocrine therapy15 (16,7%) None49 (54,4%) Histological tumor type IC NST42 (46,7%) IC NST + DCIS27 (30,0%) IC NST + LCIS4 (4,4%) ILC12 (13,3%) ILC + LCIS0 ILC + DCIS2 (2,2%) DCIS2 (2,2%) LCIS1 (1,1%) DCIS + LCIS0IC NST, invasive carcinoma of no special type; DCIS, ductal carcinoma in situ; ILC, invasive lobular carcinoma; and LCIS, lobular carcinoma in situ Table 1Characteristics of patient cohort and measurement locationsEx vivo lumpectomy specimens (N = 90)Age (years) (median, SD)56,6 (11,7)Lesion diameter (mm) (median, SD)22,6 (16,1) Neoadjuvant treatment Chemotherapy26 (28,9%) Endocrine therapy15 (16,7%) None49 (54,4%) Histological tumor type IC NST42 (46,7%) IC NST + DCIS27 (30,0%) IC NST + LCIS4 (4,4%) ILC12 (13,3%) ILC + LCIS0 ILC + DCIS2 (2,2%) DCIS2 (2,2%) LCIS1 (1,1%) DCIS + LCIS0IC NST, invasive carcinoma of no special type; DCIS, ductal carcinoma in situ; ILC, invasive lobular carcinoma; and LCIS, lobular carcinoma in situ Characteristics of patient cohort and measurement locations IC NST, invasive carcinoma of no special type; DCIS, ductal carcinoma in situ; ILC, invasive lobular carcinoma; and LCIS, lobular carcinoma in situ Classification performance Figure 8 summarizes the performance of SVM, KNN, ENS, and NN classifiers across all tumor thresholds using accuracy, AUC, sensitivity, and specificity. The ENS model demonstrated the most balanced and consistent performance, achieving high accuracy, AUC, and specificity with low variabil- ity. While KNN achieved the highest sensitivity, it showed reduced specificity and greater variability, making it less reliable overall. SVM and NN underperformed relative to ENS across most metrics. Given the context of tumor detection, where minimizing false negatives is critical, sensitivity takes precedence over specificity. The sensitivity of the ENS model can be further improved by adjusting its decision threshold, even if it results in a slight reduction in specificity. This trade-off is clinically acceptable when the priority is to ensure that tumors are not missed, and further supports the use of the ENS model as the most effective and reliable approach in this setting. Detailed performance results for the best-performing model (Ensemble), including sensitivity, specificity, ACC, and MCC with corresponding standard deviations (SD), are shown in Table 2 for each tumor threshold (TH). The ROC curves and AUCs for all tumor thresholds, based on the Ensemble model, are shown in Fig. 9. The vertical dashed line indicates the point where the curves intersect at a selected fixed specificity of 0.80, with a zoomed-in section for more details. Fig. 8Performance of four models—SVM (blue), KNN (green), ENS (yellow), and NN (pink) across accuracy (a), AUC (b), sensitivity (c), and specificity (d), averaged over all tumor thresholds. Boxes show interquartile range, with medians marked; whiskers indicate the full range Performance of four models—SVM (blue), KNN (green), ENS (yellow), and NN (pink) across accuracy (a), AUC (b), sensitivity (c), and specificity (d), averaged over all tumor thresholds. Boxes show interquartile range, with medians marked; whiskers indicate the full range Fig. 9ROC-curves for all explored tumor thresholds of the Ensemble model, including the Area Under the Curve (AUC). The vertical dashed lin1e1represents the intersection of ROC curves with a fixed specificity of 0.80, and is zoomed in for more detail ROC-curves for all explored tumor thresholds of the Ensemble model, including the Area Under the Curve (AUC). The vertical dashed lin1e1represents the intersection of ROC curves with a fixed specificity of 0.80, and is zoomed in for more detail Table 2Classification performance including standard deviation (SD) of the ensemble model for four different tumor thresholds (TH) at optimal cut-off point for sensitivity and specificity5% TH10% TH15% TH20% THSensitivity (%) (SD)0.82 (0.02)0.84 (0.01)0.85 (0.01)0.86 (0.03)Specificity (%) (SD)0.81 (0.01)0.87 (0.02)0.84 (0.02)0.86 (0.03)ACC (%) (SD)0.82 (0.01)0.85 (0.01)0.85 (0.01)0.86 (0.01)MCC (SD)0.64 (0.02)0.70 (0.03)0.69 (0.02)0.72 (0.02) Table 2Classification performance including standard deviation (SD) of the ensemble model for four different tumor thresholds (TH) at optimal cut-off point for sensitivity and specificity5% TH10% TH15% TH20% THSensitivity (%) (SD)0.82 (0.02)0.84 (0.01)0.85 (0.01)0.86 (0.03)Specificity (%) (SD)0.81 (0.01)0.87 (0.02)0.84 (0.02)0.86 (0.03)ACC (%) (SD)0.82 (0.01)0.85 (0.01)0.85 (0.01)0.86 (0.01)MCC (SD)0.64 (0.02)0.70 (0.03)0.69 (0.02)0.72 (0.02) Classification performance including standard deviation (SD) of the ensemble model for four different tumor thresholds (TH) at optimal cut-off point for sensitivity and specificity Discussion The quality of breast cancer diagnosis depends on the accuracy of biopsy techniques, as early and precise detection directly impacts treatment decisions and patient outcomes. While VABB offers advantages over CNB, it still faces challenges such as sampling errors, which can lead to diagnostic delays and increased patient burden. This study explored whether introducing DRS into VABB needles enables accurate classification of breast tissue types, including IC and DCIS, while also assessing the influence of metallic components of the needle on the optical signal. Label extraction was optimized to improve tissue classification at larger fiber distances, and various machine learning models were evaluated using optical data from ex vivo breast specimens. Incorporating DRS into the biopsy procedure provides real-time feedback on tissue composition, supporting more precise needle placement and targeted sampling. This approach can potentially improve diagnostic accuracy, reduce repeat biopsies, and minimize patient discomfort by only sampling suspicious tissue and sparing healthy areas. The findings of this study highlight several key observations. Figure 5 indicates that metal com- ponents significantly affect the optical signal, causing intensity variations due to differing reflection and scattering. SNV normalization effectively reduces this variability, aligning metal-affected and reference spectra more closely, and appears promising for minimizing metal-induced distortions in DRS measurements. For tissue classification with larger fiber distances, larger label volumes provide more comprehensive tissue assessment and improve the distinction between malignant and healthy tissue (Fig. 6). The 9 mm ellipsoid size proved optimal, maximizing specificity and AUC with- out compromising sensitivity. The Ensemble (ENS) model outperforms SVM, KNN, and NN across accuracy, AUC, and specificity, making it the best choice for classifying breast tissue in this setting (Fig. 8). When using the ENS model and a 20% tumor threshold, a sensitivity of 94% (with a specificity of 80%) is reached (Fig. 9). This tumor threshold creates a balance where the model remains sensitive to heterogeneity while still differentiating between predominantly healthy and pre- dominantly tumor samples. Overall, the Ensemble model is also more stable for all explored tumor thresholds (5–15%), whereas the other models show greater diversity. The ability to accurately clas- sify small percentages of tumor tissue, IC and/or DCIS, is beneficial when guiding biopsy needles in areas where tumor tissue may be limited. While previous studies have investigated the potential of DRS for cancer diagnostics, most have focused on free-space probes in non-metallic environments with relatively short fiber distances [16, 28–31], compared to our case, where DRS is integrated into a biopsy device that requires larger fiber separations for practical implementation. While smaller fiber distances are generally assumed to provide higher accuracy in tissue discrimination, this study demonstrates that larger fiber distances can achieve comparable classification performance. Regarding any artifacts in the spectra caused by metal, we demonstrated that SNV improved consistency among metal-affected spectra. Thereby, bridging that gap by embedding DRS around the aperture of a biopsy needle, enabling real-time tissue characterization at the point of sampling, compared to fibers implemented at the tip of the needle [21, 32]. The findings of this study may also be valuable for other fields where DRS is incorporated into surgical tools, especially in situations where metal components are present, and fiber placement is restricted to longer lengths. This study also has several limitations that must be acknowledged. Throughout this study, we carefully correlated the measured tissue locations with the corresponding regions in the H&E-stained histology sections to ensure accurate labeling of spectral data. However, it is important to recog- nize that the histology images are limited to 2D cross-sections, which represent, in our case, only the surface of the measured tissue. In contrast, the optical probe interrogates a larger 3D tissue volume, capturing spectral information from more layers and surrounding structures. As a result, key parameters such as tissue type label percentages derived from the 2D H&E sections may not fully correspond to the actual distribution of tissue components within the probed volume. This dis- crepancy could introduce variability in the interpretation of spectral data and may influence model performance compared to the ultimate clinical setting. The findings of this study confirm that SNV normalization effectively minimized, though did not completely eliminate, differences between metal- affected and reference measurements. Given the challenges of developing an algorithm in a biopsy setting with metallic components, these residual effects should be further refined during future in vivo validation to ensure reliable clinical performance. Although this study was limited to ex vivo specimens, the results provide a solid foundation for in vivo validation and further clinical translation, supporting the feasibility of implementing real-time optical feedback during VABB procedures. Previous work by de Boer et al. demonstrated that DRS spectra in breast tissue do not differ significantly between in vivo and ex vivo conditions, as the key discriminative features lie in the near-infrared wavelength range, which is minimally influenced by blood absorption and perfusion dynamics [33]. This suggests that the ex vivo findings presented here are representative of in vivo conditions. Although the study included 90 patients and over 500 measurements, the limited number of DCIS and LCIS cases may affect generalizability across breast cancer subtypes. Future studies with more balanced cohorts and repeated measurements are needed to evaluate intra-patient variability and ensure robust performance across patient-specific tissue heterogeneity. Future in vivo studies should confirm the reliability of DRS measurements and assess the potential effects of patient movement. Clinical implementation will further require compatibility with existing VABB systems, simplified calibration, real-time feedback of optical measurements, and validation of safety and sterility for regulatory approval. While the models were optimized and validated using repeated cross-validation, future work should include independent external datasets to confirm generalizability. To summarize, this study demonstrates the potential of integrating DRS into larger biopsy needles to enable real-time classification of tumor tissue (IC and/or DCIS), thereby improving diagnostic accuracy, reducing repeat procedures, and enhancing clinical decision-making. This approach offers a more precise and efficient alternative to conventional biopsy techniques, with clear benefits for both patients and healthcare providers. While promising, further work is needed to address technical challenges and validate performance in clinical settings. With continued advancements in optical sensing and machine learning–based tissue classification, breast cancer diagnostics may shift toward more data-driven, minimally invasive, and highly accurate approaches. Conclusion This study demonstrates that Diffuse Reflectance Spectroscopy can be successfully incorporated into vacuum-assisted breast biopsy procedures to improve tumor detection. By addressing metal-induced spectral interference and optimizing label extraction for increased fiber distances, we achieved robust tumor classification using machine learning, with ensemble models reaching sensitivities of up to 94% at a 20% tumor threshold. Importantly, the model was also capable of detecting IC and DCIS even at lower tumor thresholds, highlighting its potential for identifying small or heterogeneous tumor areas. Future in vivo studies are essential to confirm these results under clinical conditions and to refine the approach for routine implementation. This approach offers a more precise alternative to conventional biopsy, with clear potential to improve diagnostic accuracy and patient care in breast cancer diagnostics.