Emerging and future use of intra-surgical volumetric X-ray imaging and adjuvant tools for decision support in breast-conserving surgery

Abstract

Breast-conserving surgery requires that resection margins be cancer-free, but re-excision rates due to positive margins have remained near 20% for much of the last decade with high variability between surgical centers. Recent studies have demonstrated that volumetric X-ray imaging improves margin assessment over standard techniques, given the speed of image reconstruction and full three-dimensional sensing of all margins. Deep learning approaches for automated analysis of volumetric medical image data are gaining traction and could play an important role streamlining the clinical workflow for intra-surgical specimen imaging. X-ray imaging systems currently deployed in clinical studies suffer from poor tumor-to-fibroglandular tissue contrast, motivating the development of adjuvant tools that could potentially complement volumetric X-ray scanning and further improve the future of intra-surgical margin assessment by real-time augmented guidance for the surgeon.

Graphical Abstract

1. Introduction

1.1. Importance and scope

In 2021, breast cancer became the most common cancer globally (excluding non-melanoma skin cancer), accounting for 12% of all new annual cancer diagnoses worldwide [1]. Breast-conserving surgery (BCS) combined with radiotherapy achieves equivalent disease-free survival as mastectomy for early-stage, localized breast cancer and is therefore the treatment of choice for early-stage disease [2]. Successful treatment requires that resection margins be free of malignancy, and Society of Surgical Oncology–American Society for Radiation Oncology (SSO-ASTRO) consensus guidelines published in 2014 [3] and 2016 [4] have led to a statistically significant reduction in re-excision rates [5]. Nevertheless, re-excision rates due to positive margins have remained near 20% for much of the last decade [6,7]. Re-excision procedures pose substantial health and financial burdens to patients and the healthcare system [3,8], and so a need exists for improving intra-surgical decision-making tools for margin assessment, such that positive margin rates, and thus re-excision procedures, become less frequent.

This article focuses on advancements in intra-surgical imaging guidance for BCS margin assessment over the last five years. Emphasis is placed on X-ray imaging technologies for the following reasons: X-ray imaging is ubiquitous in clinical diagnostic imaging; radiologists are already experts at interpreting images derived from X-ray attenuation; intra-surgical specimen radiography (SR) (i.e., two-dimensional projection X-ray imaging) is one of the most common techniques for BCS margin assessment today [9]; and finally, recent pre-clinical and clinical studies focused on X-ray volumetric specimen imaging (VSI) show promise for improving intra-surgical BCS margin assessment [10–21]. In Section 1.2, the state-of-the-art for BCS margin assessment is briefly reviewed. Section 2 highlights key studies that evaluated the efficacy of VSI using X-rays for BCS margin assessment. In Section 3, phase contrast X-ray imaging (Section 3.1) and a range of adjuvant tools – particularly optical sensing – (Section 3.2) are discussed for overcoming the limited soft tissue contrast provided by current X-ray specimen imaging systems. In Section 4, deep learning approaches in X-ray VSI for decision support and radiological interpretation guidance are discussed. Section 5 presents concluding remarks and future directions for X-ray VSI for margin assessment in BCS.

1.2. State-of-the-art

Current state-of-the-art methods for intra-surgical margin assessment in BCS include visual inspection and palpation by the surgeon, meticulous recording of resection anatomical orientation, SR, and at some medical centers, routine circumferential cavity shaving and intraoperative pathological processing (i.e., frozen section pathology and touch prep cytology) [22]. Although intraoperative pathological processing provides high diagnostic accuracy, these methods are resource intensive and suffer slow turnaround times, limiting their widespread adoption [23]. In the case of SR, the tissue is imaged immediately after resection, and within minutes, a radiologist interprets the radiograph and notifies the surgeon of suspicious peripheral margins revealed in the two-dimensional image. The margins normal to the imaging axis – typically the superficial and deep (i.e., anterior and posterior) margins – are occluded in the projection and cannot be assessed. Furthermore, the predictive value of SR is low for ductal carcinoma in situ (DCIS), pre-invasive cancer that is frequently nonpalpable, discontinuous, and responsible for one of the largest shares of positive margins [7]. The specimen subsequently undergoes post-operative pathological processing to determine final margin status, a process that takes several hours to days to complete. Re-excision BCS procedures – based on post-operative pathologic final margin status – remain high despite these standard-of-care measures.

2. Trend toward X-ray micro-computed tomography and tomosynthesis

Intra-surgical SR and radiological interpretation are already commonplace in standard-of-care BCS. Thus, adopting VSI in the clinic appears feasible without major changes in surgical workflow. Several recent studies have demonstrated the benefits of VSI over conventional SR for BCS margin assessment (Figure 1i) and tested the feasibility of translating VSI into the clinic. For this article, intra-operative imaging refers to the process of image acquisition, reconstruction, and interpretation, and in general, these steps should take 20 minutes or less to be considered clinically feasible. Some studies focused on micro-computed tomography (micro-CT), which involves the acquisition of X-ray projection images over a full 360° around the target and filtered back projection to reconstruct the scan volume at high spatial resolution. Other studies used tomosynthesis, which involves acquisition of X-ray projection images over a narrow range of angles typically less than 60°, followed by volume reconstruction into axial planes. The benefit of tomosynthesis over micro-CT is primarily increased acquisition and/or reconstruction speed at the expense of lower reconstruction quality; the drawback is significant blurring between planes of reconstruction. The following paragraphs highlight results from a subset of clinical studies over the last five years focused on VSI for BCS margin assessment. Salient findings from these studies are summarized in Table 1.

Figure 1.

Recent studies focused on micro-CT volumetric specimen imaging for breast-conserving surgery (BCS) margin guidance. (i) Kulkarni et al. demonstrated that positive margins missed by multi-view projection radiography could be identified by micro-CT scanning – adapted from [21]. (ii) Janssen et al. developed a custom specimen container for localizing suspected margins prior to pathological processing – adapted from Ref [15]. (iii) McClatchy et al. demonstrated that X-ray attenuation lacks contrast between non-malignant fibroglandular tissue and malignancy – adapted from [13]. (iv) DiCorpo et al. demonstrated proof-of-principle tumor segmentation based on X-ray attenuation thresholding – adapted from [18].

Table 1.

Summary of recent volumetric X-ray specimen imaging studies for intra-surgical margin assessment during breast-conserving surgery. Unless otherwise specified, all measurements/results focused on final margin status on the specimen level and were assessed relative to final margin status determined by post-operative pathological analysis. Bold numbers in the Results column correspond to the bolded modality in the Measurements column.

Reference	Specimen Count	Study Type	Measurement(s)	Result(s)		Positive Margin Basis for Comparison to Pathology	(A)cquisition, (R)econstruction, (I)nterpretation Time		Imaging Systema
Tang et al. 2016 [10]	50	Observational; single-center	Micro-CT vs. SR; largest tumor dimension	Pearson’s correlation coefficient of 0.82 vs. 0.40		Positive margins not assessed	A R I	7 min 7 min Not reported	Bruker SkyScan 1173
Urano et al. 2016 [11]	65	Observational; single-center	Tomosynthesis vs. SR; extent of invasive lesion detected by 1 blinded radiologist	Clear whole lesion delineation in 45% vs. 6% of specimen images		Positive margins not assessed	None reported		Siemens MAMMOMAT Inspiration
Amer et al. 2017 [12]	102	Observational; single-center	Tomosynthesis vs. SR independently analyzed by 2 blinded radiologists; direction closest to margin and margin width	Overall accuracy of 69% vs. 40%		<1 mm for invasive cancers; 5 mm for DCIS	None reported		Siemens MAMMOMAT Inspiration
McClatchy et al. 2018 [13]	32	Observational; single-center	Micro-CT read by 1 blinded radiologist	Accuracy (64%), sensitivity (50%), specificity (67%), NPV (89%), PPV (20%)		Tumor on margin	A R I	1 min 2 min Not reported	PerkinElmer IVIS SpectrumCT
Qiu et al. 2018 [14]	30	Observational; single-center	Micro-CT read jointly by 2 blinded radiologists	Accuracy (86%), sensitivity (56%), specificity (100%), NPV (83%), PPV (100%)		Tumor on margin for invasive cancers; <2 mm for DCIS	A R I	7 min 1.5 min 5 min	Bruker SkyScan 1275
Janssen et al. 2019 [15]	100	Observational; single-center	Micro-CT independently read by 4 blinded observers in 2 phases (P1-P2), read by 1 blinded observer in phase 3 (P3)b	Accuracy	(P1: 63%; P2: 72%; P3: 70%),	Tumor on margin	A R A+R+I	8–10 min 4–7 min 20 min	Bruker Skyscan 1275B
				sensitivity	(P1: 38%; P2: 40%; P3: 38%),
				specificity	(P1: 70%; P2: 78%; P3: 78%),
				PPV	(P1: 22%; P2: 26%; P3: 30%),
				κ	(P1: 0.31; P2: 0.23)
Park et al. 2019 [16]	99	Observational; single-center	Tomosynthesis read by 1 blinded radiologist vs. “SEP” including SR	Sensitivity	(74% vs. 84%),	Tumor on margin	A R I	1 min Not reported Not reported	Kubtec Medical Imaging Mozart
				specificity	(91% vs. 78%),
				NPV	(99% vs. 99%),
				PPV	(21.5% vs. 11%)
Garlaschi et al. 2019 [17]	89	Interventional; single-center	Tomosynthesis independently read by 2 blinded radiologists with access to SR	ROC curve AUC of 0.82 vs. 0.65		<1 mm	None reported		Kubtec Medical Imaging Mozart
DiCorpo et al. 2020 [18]	173	Observational; single-center	Micro-CT read by more than one reader	Accuracy of 93% detecting cases with positive margins		Tumor on margin	A+R I	8–10 min “Hours”	Bruker SkyScan 1173, 1275 and Nikon XTH225
Partain et al. 2020 [19]	191	Interventional; two centers	Tomosynthesis (191 specimens, site 1) vs. SR (466 specimens, sites 1–2) read intraoperatively by surgeons to guide targeted cavity shaves	Final positive margin rate of 4% vs 9%		Tumor on margin	None reported		Kubtec Medical Imaging Mozart
Romanucci et al. 2021 [20]	170	Observational; single-center	Tomosynthesis vs. SR independently read by 2 blinded radiologists; tumor-free resection margin distances	Pearson’s correlation coefficient of 0.92 vs. 0.79 in cranial-caudal view and 0.92 vs. 0.72 in medial- lateral view	Positive margins not assessed	None reported		Hologic Selenia Dimensions
Kulkarni et al. 2021 [21]	200	Observational; two centers	Micro-CT vs. tomosynthesis vs. SR independently read by 3 blinded radiologists at 2 sites (S1-S2)c	Sensitivity	(91–94% vs. 38–53% vs. 34–49%),	Tumor on margin for invasive cancers; <2 mm for DCIS	A+R I	5–7 min 1–2 min	Clarix Imaging Corp. “Prototype system”
				Specificity	(81–85% vs. 71–88% vs. 78–88%),
				PPV	(25–30% vs. 11–17% vs. 10–17%),
				NPV	(99% vs. 96% vs. 95–96%)

SR = specimen radiography; IDCa = invasive ductal carcinoma; ILCa = invasive lobular carcinoma; NPV = negative predictive value; PPV = positive predictive value; ROC = receiver operating characteristic; AUC = area under the curve; SEP = standard extensive processing, unique to Ref [16]; κ = Cohen’s kappa for interobserver agreement.

^aCompanies and locations: Bruker, Kontich, Belgium; Siemens, Munich, Germany; PerkinElmer, Hopkinton, MA, USA; Kubtec Medical Imaging, Stratford, CT, USA; Nikon, Tokyo, Japan; Hologic, Bedford, MA, USA; Clarix Imaging Corp., Chicago, IL, USA.

^bResults are given as averages from the 4 blinded observers.

^cResults reported on an individual margin level (i.e., six margins per specimen were individually tallied).

In 2016, Tang et al. demonstrated that micro-CT scans of 50 BCS resections provided more accurate primary tumor dimension than SR when compared to gold standard pathologic tumor dimension (Pearson’s correlation coefficient of 0.82 vs. 0.40) [10]. In 2018, McClatchy et al. correlated readings of 32 whole BCS resection micro-CT scans by a breast radiologist to final histopathology, finding that micro-CT matched final pathological diagnosis of margins in 64% of cases and provided a negative predictive value of 89% (positive margin definition: tumor on the specimen edge) [13]. The study also demonstrated full three-dimensional scanning of whole resections in a clinically relevant timeframe: less than four minutes total for image acquisition and reconstruction. Another study in 2018 by Qiu et al. assessed the feasibility of micro-CT for intraoperative BCS margin assessment by scanning 30 consecutive whole resections and having two investigators interpret the volumetric scans together in 15 minutes or less [14]. The accuracy, sensitivity, and specificity of the micro-CT-based readings were 86%, 56%, and 100%, respectively (positive margin definition: invasive tumor on the specimen edge; DCIS within 2 mm of specimen edge). In 2019, Janssen et al. reported results from a multi-phase study in which four observers retrospectively analyzed two sets of 30 whole resection micro-CT scans and then prospectively analyzed a final set of 40 whole resection scans [15]. The prospective study resulted in an overall accuracy of 70% classifying margins when compared to final pathology (positive margin definition: tumor on the specimen edge). However, the authors found only fair interobserver agreement (Cohen’s kappa of 0.31 and 0.23 in the first and second retrospective studies, respectively). Park et al. published a 2019 study focused on the ability of tomosynthesis to detect positive margins in 99 whole BCS resections relative to standard-of-care intraoperative “extensive processing” that utilized SR [16]. The study concluded that tomosynthesis performed similarly to labor-intensive standard-of-care processing, yielding sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of 74%, 91% 22%, and 99%, respectively (positive margin definition: tumor on the specimen edge).

Studies through 2019 offered valuable insight into the feasibility and the potential value of using X-ray VSI for BCS margin assessment. However, relatively small sample sizes limited the generalizability of findings. In 2020, DiCorpo et al. published a study in which 173 whole BCS resections were imaged with micro-CT for intra-surgical margin assessment, achieving overall positive margin accuracy of 93% (positive margin definition: tumor on the specimen edge). The study concluded that micro-CT could detect malignancy on small regions of the margin likely to be missed by pathological sectioning [18]. In 2021, Kulkarni et al. published a two-center study in which 200 cases were enrolled, and three radiologists interpreted SR, tomosynthesis, and micro-CT VSI of each specimen [21]. The three radiologists reported AUC values of 0.91, 0.90, and 0.94 using VSI, showing relative improvement over the AUC values using SR by 54%, 13%, and 40% and tomosynthesis by 32% and 11% (positive margin definition: invasive tumor on the specimen edge; DCIS within 2 mm of specimen edge).

X-ray VSI provides improved sensing of BCS resection margins relative to standard-of-care margin assessment tools (Table 1). Gold standard, post-operative histopathological margin analysis involves gross dissection (“bread loafing” or slicing the specimen into millimeters thick sections), and then microscopic evaluation is selectively performed on thin (~5 μm) slices of tissue sections. Thus, gold standard histopathology inherently involves substantial under sampling of the margin. For this reason, it is possible that high resolution X-ray VSI may detect residual disease on the margin that is not detected by gross dissection or histopathological analysis [13,18]. At least one innovative method has been developed to facilitate exact VSI correlation to pathology (Figure 1ii), and such solutions should be prioritized in future BCS margin assessment studies to meaningfully gauge efficacy. X-ray VSI provides superior correlation with pathology with respect to largest tumor dimension [10], lesion delineation [11], and tumor-free margin distances [20] compared to SR. Notably, X-ray VSI suffers from low PPV, typically ranging from 20–30% [13,15,16,21]. This trend – the result of a high number of false positive readings (Figure 1iii) – is discussed in more detail in Section 3.1 X-ray VSI acquisition and reconstruction can be achieved in less than 15 minutes [10,13,14,18,21], demonstrating that the technology is appropriate for intra-surgical use. While fewer studies report the time required for scan interpretation, evidence suggests that a trained observer can interpret a volumetric scan of a BCS resection in five minutes or less [14,21]. Nevertheless, developing volumetric analysis tools to guide tumor segmentation (Figure 1iv) and suspicious margin identification has garnered significant interest in medical imaging in recent years and will likely empower clinicians with improved BCS margin assessment in years to come (see Section 4). It is noteworthy that most VSI studies in the last five years were observational in nature – not in any way connected to patient care or clinical decision making. However, two studies involved intra-surgical interpretation of VSI scans to guide resection [17,19]. More interventional, prospective studies are required to fully capture the benefits of X-ray VSI relative to standard-of-care techniques for BCS margin assessment.

Finally, the detection of pre-invasive DCIS at or near the margin poses a unique challenge. The SSO-ASTRO consensus guidelines published in 2016 state that for specimens containing only DCIS, tumor within 2 mm of the margin should be considered positive, while specimens containing both pre-invasive and invasive components, the standard “tumor on ink” (i.e., tumor on the margin) rule defines a positive margin [4]. Most DCIS lesions are diagnosed by X-ray mammography through the identification of specific types and structures of microcalcification deposits [24]. It is possible that high resolution X-ray VSI could leverage microcalcification signatures to differentiate DCIS and benign lesions, although to the authors’ knowledge, this has not yet been demonstrated.

3. The problem of limited soft tissue contrast

3.1. Phase-contrast computed tomography

It is well known that X-ray imaging suffers from poor contrast between soft tissues of similar density. In the context of imaging BCS resections, this limitation impacts the ability to differentiate solid tumor and non-malignant fibroglandular tissues [10,13,14,18,25]. If scans are interpreted conservatively, this limitation results in a high number of false positives, leading to unnecessary removal of additional tissue from the surgical cavity and prolonged procedural times. At least one potential solution exists for overcoming this limitation of current X-ray VSI systems. Massimi et al. recently published an inceptive study using X-ray phase contrast CT for detecting involved margins in breast specimens [26]. They demonstrated superior sensitivity and comparable specificity to conventional SR for detecting disease at the margin, and importantly, provided an example of heterogeneity within a solid tumor mass that correlated strongly with microscopic histological analysis (Figure 2i). While the study provides a valuable example of technological advances in micro-CT specimen scanning, additional analysis using X-ray phase contrast CT with exact correlations to histology are required to demonstrate sensitivity between healthy breast parenchyma and malignant tissues.

Figure 2.

Recent advances in intra-surgical margin assessment technologies. (i) X-ray phase contrast CT provides increased contrast to soft tissue in breast specimens – adapted from Ref [26]. (ii) Examples of emerging techniques that could function as adjuvant tools alongside X-ray volumetric specimen imaging: intra-surgical pathology – adapted from Ref [27]; Cerenkov luminescence – adapted from Ref [32]; Raman spectroscopy – adapted from Ref [31]; bioimpedance spectroscopy – adapted from Ref [29]; radiofrequency spectroscopy – adapted from Ref [28]; Lumicell (LUM015) fluorescence – adapted from Ref [33]; optical spectroscopy – adapted from Ref [30]; and 5-aminolevulinic acid hydrochloride (5-ALA HCl) fluorescence – adapted from Ref [34].

3.2. Adjuvant tools

A large number of alternative technologies have been investigated for BCS margin assessment in recent years, spanning intra-surgical pathological processing [27], radiofrequency spectroscopy [28], bioimpedance spectroscopy [29], and optical methods, just to name a few (Figure 2ii). Optical methods in particular span a wide range of implementations, formfactors, and mechanisms of contrast, including diffuse spectroscopy diffuse spectroscopy [30], Raman spectroscopy [31], Cerenkov luminescence [32], and various fluorescent probes [33,34]. These methods have demonstrated strong potential for the identification of malignancy and even specific tissue subtypes on the margin. Multimodal optical imaging has also recently been proposed for margin assessment. For example, tissue auto-fluorescence combined with Raman spectroscopy demonstrated >99% sensitivity to the challenging pre-invasive tissue subtype DCIS [35].

Each proposed technique presents pros and cons. An extensive review of these margin assessment techniques is beyond the scope of this article, and the interested reader is directed to a recent review focused on the topic [36]. While one or more individual margin assessment technologies may successfully translate into the clinic, it is possible that multiple solutions working in consort will provide optimal performance. These novel methods may function as adjuvant tools for BCS margin assessment alongside X-ray VSI in the future, providing the much-needed sensitivity to differentiate non-malignant and malignant fibrous tissues on the margin, while benefitting from the robust three-dimensional morphological sensing of X-ray VSI.

4. Intra-surgical analysis by deep learning approaches

One of the key problems in using advanced imaging tools during a surgical procedure is that the volume and complexity of information generated can be too much for reasonable decision making by the surgeon or consulting radiologist and pathologist. Multi-scale datasets that have combinations of macroscopic, radiologic, and microscopic information present the problem of how to distill this information for fast decision making. In the past five years, deep learning (DL) approaches have garnered significant attention in the application of digital mammography [37,38], and there is a rapidly growing body of work with conventional CT datasets which can, in principle, also be applicable to the micro-CT domain. This section highlights recent DL studies that can improve X-ray VSI in three important areas: 1) high-quality image reconstruction, 2) automated segmentation of tumor margins (Figure 3i–ii), and 3) multi-modal fusion of imaging data with adjuvant tools and clinical records (Figure 3iii–iv). As this review is application focused, many of the technical aspects of DL are not covered in detail, but the interested reader is referred to additional topical reviews for DL in medical imaging applications [39–41].

Figure 3.

Potential for automated analysis of breast specimen images using deep learning. (i) Diagram of deep learning-based segmentation models for 2D and 3D margin analysis – adapted from [49]. (ii) Deep learning-based tumor segmentation in lung specimens with visual overlay of model predictions by Moriya et al. [52]. (iii) Diagram of deep learning-based data fusion of clinical records and adjuvant tools to improve segmentation and overall risk stratification. (iv) Example of multimodal data fusion between medical CT imaging and electronic medical record (EMR) data by Huang et al. [53].

Radial sampling with volumetric reconstruction is fundamental to all CT methods, and involves practical trade-offs in image quality, radiation dose, and acquisition speed. In the context of ex vivo specimen imaging, radiation dose to the specimen is immaterial; however, acquisition time remains an important constraint. Multiple studies have demonstrated that DL models can be trained to approximate high-quality reconstructions (sampled at 0.5° or 1° intervals) using very sparse imaging data (8–12 times fewer projections) as input [42,43]. These sparse sampling schemes could significantly reduce acquisition time and provide meaningful time/cost savings in surgical settings. Importantly, there is ongoing work to investigate whether or not DL methods are fundamentally solving the sparse-view CT inverse problem (similar to compressive sensing techniques) as well as to make datasets to evaluate these methods publicly available [44]. Beyond sparse reconstruction, DL has also been utilized for noise reduction and metal artifact removal [45,46], which could potentially improve the overall quality of the volumetric data in preparation for subsequent margin analysis.

For volumetric imaging to be utilized for decision support, careful review of the full specimen volume is necessary to delineate normal, abnormal benign, and suspected malignant tissues extending from the primary tumor volume toward the margin. In recent years, DL methods for automated segmentation of 3D medical image datasets have advanced substantially [47]. Multiple studies have now demonstrated promising results using DL in resected specimen margin delineation [48–50]. Figure 3i shows examples of both 2D and 3D segmentation networks which take specimen images as input and output the tumor segmentation. Figure 3ii demonstrates how these segmentation networks can extend to multi-class labeling schemes as well as provide heatmap visualizations of high-risk areas. An important limitation of these studies, however, is sample size (min-max samples: 7–24 tissue specimens; 20–21 mice). Larger scale datasets and assessments of automated segmentation of tumor volumes using micro-CT are greatly needed. Recent developments in DL (such as self-supervised learning and attention mechanisms) can help address many of the limitations of small and imperfectly labeled datasets common in biomedical image segmentation [47]. These approaches should be highly beneficial for future development of micro-CT segmentation models.

Lastly, the integration of additional data sources from adjuvant tools and/or pre-operative biopsy histology represent an emerging opportunity to provide more robust risk stratification and probability estimates throughout the micro-CT volume. In the case of adjuvant tools, ideally this could be achieved spatially correlating additional measurements with the surface mesh of the micro-CT volume. DL methods which leverage multi-modal image volumes can outperform single modality networks on medical image segmentation [51]. Additionally, because DL is agnostic to input data structures, clinical data (e.g., pre-operative cancer diagnoses, tumor genomics) can also be used as input to separate branches of a multi-modal classification network with a separate feature encoder for each data source (Figure 3iii). Huang et al. demonstrated the utility of a multi-modal DL approach for detection of pulmonary emboli in CT scans (Figure 3iv). Their work provides a proof-of-principle for future studies that integrate both clinical and imaging data in a unified probabilistic model.

In summary, streamlining of micro-CT acquisition and analysis will be crucial for broad adoption in the surgical setting. DL methods are becoming increasingly capable in several foundational medical imaging tasks and can add value to multiple aspects of X-ray VSI. Because application of DL within VSI is still relatively new, an urgent need exists for open datasets that serve as benchmarks to foster further research and development of DL in VSI.

5. Conclusions

Volumetric X-ray scanning of BCS specimens shows strong potential for improving breast-conserving surgery margin assessment, although X-ray imaging alone lacks contrast between fibrous breast tissues, suggesting value in multimodal solutions using adjuvant tools. Computational decision support in medical imaging in general is rapidly advancing but is still needed in BCS margin assessment. Collaboration between surgeons, radiologists, pathologists, and engineers is needed to validate proposed techniques with exact correlation to gold standard pathological results. Emphasis should be placed on making image data and decision support model output interpretable and intuitive for use by clinicians (e.g., model output is mapped to actual tissue surfaces). Finally, a need exists for larger multi-center, prospective studies to fully capture the benefits of X-ray VSI and to determine if these technologies can be translated into standard BCS clinical practice.

Highlights.

• Emerging volumetric X-ray imaging technologies show strong potential for improving breast-conserving surgery (BCS) margin assessment.

• X-ray imaging alone lacks contrast between fibrous breast tissues, suggesting value in multimodal solutions using adjuvant tools.

• Advances in computational decision support and clinical interaction with and interpretation of volumetric data are needed.

• Collaboration between surgeons, radiologists, pathologists, and engineers is necessary to validate and improve the workflow and potential integration of these emerging tools into BCS.