Novel rapid intraoperative qualitative tumor detection by a residual convolutional neural network using label-free stimulated Raman scattering microscopy
Acta Neuropathologica Communications volume 10, Article number: 109 (2022)
Determining the presence of tumor in biopsies and the decision-making during resections is often dependent on intraoperative rapid frozen-section histopathology. Recently, stimulated Raman scattering microscopy has been introduced to rapidly generate digital hematoxylin-and-eosin-stained-like images (stimulated Raman histology) for intraoperative analysis. To enable intraoperative prediction of tumor presence, we aimed to develop a new deep residual convolutional neural network in an automated pipeline and tested its validity. In a monocentric prospective clinical study with 94 patients undergoing biopsy, brain or spinal tumor resection, Stimulated Raman histology images of intraoperative tissue samples were obtained using a fiber-laser-based stimulated Raman scattering microscope. A residual network was established and trained in ResNetV50 to predict three classes for each image: (1) tumor, (2) non-tumor, and (3) low-quality. The residual network was validated on images obtained in three small random areas within the tissue samples and were blindly independently reviewed by a neuropathologist as ground truth. 402 images derived from 132 tissue samples were analyzed representing the entire spectrum of neurooncological surgery. The automated workflow took in a mean of 240 s per case, and the residual network correctly classified tumor (305/326), non-tumorous tissue (49/67), and low-quality (6/9) images with an inter-rater agreement of 89.6% (κ = 0.671). An excellent internal consistency was found among the random areas with 90.2% (Cα = 0.942) accuracy. In conclusion, the novel stimulated Raman histology-based residual network can reliably detect the microscopic presence of tumor and differentiate from non-tumorous brain tissue in resection and biopsy samples within 4 min and may pave a promising way for an alternative rapid intraoperative histopathological decision-making tool.
Rapid intraoperative histopathological analysis of fresh frozen tumor tissue is an essential tool for biopsy and surgical resection control . This is based on a time-consuming and costly intraoperative pathology workflow for intraoperative decision-making that often leads to prolonged surgery times with the patient kept in anesthesia waiting for the pathological diagnosis [2, 3]. In addition, not every hospital providing surgeries in patients with brain tumors holds a (neuro)pathology unit available for rapid fresh frozen diagnostics and even if, it is usually only available during normal working hours. Recently, a fully automated label-free optical laser system creating digital hematoxylin-and-eosin (H&E)-like images called stimulated Raman histology (SRH) has been introduced as an alternative to fresh frozen conventional histopathological section diagnostics [4,5,6,7,8]. In brief, SRH is an optical digital molecular microscopy technique that exploits the intrinsic optical properties of biological macromolecules, such as lipids, proteins, and ribonucleic acids, and converts them into digital H&E-like images within 3–4 min. The SRH images are then to be reviewed by a neuropathologist or available for analyses of a machine learning algorithm. The combination of SRH and the use of machine-learning techniques for analysis in a semi-automated workflow has already been explored in some proof-of-concepts investigations [8, 9]. Using convolutional neural networks and the associated evaluation of SRH images, various neurosurgical tumor entities can already be reliably detected [4, 5]. This possibility provides the neurosurgeon during the surgical resection with important information about the tumor entity and glial malignancy to tailor the surgery accordingly. However, the recently introduced CNN to determine the tumor entity and malignancy in the case of gliomas is especially limited in reliably detecting tumor tissue in the tumor margins for resection extent and biopsy control. This study aimed to develop a novel SRH-coupled deep residual CNN for qualitative brain tumor detection independent of entity and dignity and test its reliability to distinguish between tumor and non-tumor brain tissue in a semi-automated clinical intraoperative workflow.
Materials and methods
In a prospective single-center study design patients were recruited in 2021 after they were indicated for surgery or stereotactic biopsy based on decisions of an interdisciplinary neuro-oncological tumor board panel. This study was reviewed and approved by the local ethics committee (Nr. 21-1238). All patients gave their written informed consent for the scientific use of their data according to European law. Inclusion criteria comprised: (1) Suspected tumor lesions of the central nervous system, (2) aged over 18 years, and (3) the patient is willing and able to give informed consent for participation. Exclusion criteria comprised: (1) Patient does not agree to participate in the study or (2) unable to or unwilling to give informed consent. Specimen samples were excluded if the collected specimen was inadequate, e.g. broken slide or specimen sample size under 1.7 mm diameter for SRH imaging. All specimen samples obtained were imaged immediately after collection by the intraoperative label-free fiber-laser-based SRS microscope and evaluated by the CNN as the output class (1) tumor, (2) non-tumor, or (3) low quality. An independent neuropathologist reviewed all generated SRH images as a control arm.
Specimen collection and intraoperative SRH imaging method
Tissue samples were intraoperatively obtained from the lesion itself in surgery and stereotactic biopsies as well as from the approach area in surgery. These samples did not interfere with or differ from the regular tumor sampling sent to (neuro)pathology to enable histopathological diagnosis. For virtual imaging of fresh specimens, a small unprocessed specimen sample (3–4 mm in size) was squeezed onto a glass slide with the help of the cover glass and imaged by a clinical, fiber-laser based stimulated Raman scattering microscope (Invenio Imaging Inc, Santa Clara, CA, USA) . In brief, four main components are used here: (1) a fiber-coupled microscope, (2) a dual-wavelength fiber-laser module, (3) a laser and microscope control module, and (4) a computer for data displaying, processing, and application of a CNN. Spectral Raman differences and concentration cause lipids and proteins to reflect differently, creating contrast. Tissue samples were imaged at two Raman shift wavenumbers, 2845 cm−1 (CH2/lipid channel) and 2940 cm−1 (CH3/protein and ribonucleic acids). Via another process, a three-channel image was created after image subtraction, resulting in a digital H&E-like stained 3.06 mm2 image for intraoperative and pathological view, which is referred to as Stimulated Raman Histology (SRH).
Machine learning algorithm
A residual convolutional neural network (CNN) architecture named ResNet50v2 was used for image classification on 300 × 300 pixel SRH image patches. The final layers in the network were modified for classification into three classes: (1) tumor, (2) non-tumor, and (3) low quality. The network was trained on a separate previously acquired and annotated dataset of 570 whole-slide SRH images representing the full spectrum of neurooncological surgery that resulted in 1.2 million labeled patches after patch extraction using 300 × 300 pixels sliding window technique with a step of 100 pixels in both x- and y-direction. The CNN training was done on 90% of the training data set, while preliminary validation of the CNN after each epoch was done on 10% of the training dataset. The imbalance of the classes was taken into account by using the inverse class frequencies as weights for the categorical cross-entropy loss of the CNN. The training was stopped after the preliminary validation accuracy exceeded 95% and the loss was lower than 0.10.
The CNN was re-trained a second and a third time with different random seeds on the same external dataset for reliability and ROC analysis (see Fig. 1).
CNN Segmentation and probability heatmap of SRH Images
To run the CNNs on whole-slide SRH images, testing patches were extracted using a 300 × 300-pixels sliding window with a step size of 300-pixels in both x- and y-direction. In this way, there were no overlapping patches in testing. Patches that had low quality as top-1 class were excluded from the analysis of the final whole-slide prediction, thus only “high-quality” patches probabilities were averaged over the whole-slide image for the final prediction. Heatmaps for whole-slide SRH images (e.g. 3600 × 3900 true image pixels) were created by interpolating the images containing predictions at the patch level (e.g. 12 × 13 patch prediction pixels containing class probabilities). The final probability heatmap of the presence of tumor (red), non-tumor (green), and low-quality (blue) is coded on top of the SRH image as a semi-transparent overlay and could assist the user in interpreting the SRH images by color-coded class probabilities (see Figs. 2, 3).
The SRH images analyzed by CNNs were reviewed as a high-resolution PNG file with 3600 × 3900 pixels by one board-certified independent neuropathologist, blinded to the CNN results and clinical data, who analyzed, and classified into the three diagnostic classes, (1) tumor, (2) non-tumor, and (3) low quality by a binary evaluation pattern. The predictions of the CNNs and the independent neuropathologist were compared by review on a single image-by-image basis. The highest probability prediction value of the corresponding output class from the CNNs was chosen and thus converted into a binary form to enable the direct comparison to the independent blinded neuropathologist’s evaluation.
Statistical analysis was performed using SPSS Statistics Version 28 (IBM, Chicago IL). For descriptive statistics, continuous values are given in mean with standard deviation, and ordinal and categorical variables are stated in counts and percentages. Internal consistency (IC) of random areas within a specimen sample was calculated using Cronbach's alpha (C α). The inter-rater reliability (three raters) of the probability values between the three trained CNNs was calculated by the intraclass correlation coefficients (ICC) (single-measurement, absolute-agreement, 2-way mixed-effects model). The inter-rater reliability (two raters) between CNNs and neuropathologist was calculated using Cohen´s Kappa (κ). The diagnostic quality of the trained and applied CNNs was validated by the Receiver operating characteristic (ROC) analysis. Values > 0.8 were considered as excellent and > 0.9 as outstanding . As a non-parametric statistic test, Kendall-W-Test was used for significance proofing between CNNs probability values. P-values below 0.05 were considered statistically significant.
Demographics and baseline characteristics
We included 94 brain tumor patients, 49 (52.1.%) of them were female. The type of procedure to obtain specimen samples comprised surgical resections in 77 (81.3%) and stereotactic biopsies in 17 (18.1%) patients.
The final histopathological diagnosis was benign to malignant gliomas in 27 (28.7%), various carcinoma and melanoma metastases in 26 (26.6%), benign to anaplastic meningioma in 13 (13.8%), pituitary adenoma in 6 (6.4%), craniopharyngioma in 2 (2.1%), ependymoma in 2 (2.1%), lymphoma in 7 (7.4%), medulloblastoma in 2 (2.1%) patients, other entities such as epidermoid-cyst, ganglioneurinoma, hemangioblastoma, lipoma, rathke cleft-cyst, schwannoma in 6 (6.4%) patients, and gliosis in 3 (3.2%) patients by stereotactic biopsy.
A total of 226 specimen samples were collected for SRH imaging (range 1–8 specimens per patient) resulting in a total of 592 SRH images that were generated (range 1–4 different SRH images per specimen sample). A minimum of three different random areas (A, B, C) per slide were scanned in 132 specimens (58.4%) resulting in 402 SRH images in total for statistical analyses. In three cases of the 402 SRH images, one of the three random areas showed a 100% probability value of low quality, so a fourth random area (D) was obtained for SRH image analysis.
Seventeen (7.5%) microscopic specimen samples resulting in 34 (5.7%) SRH images were obtained as negative controls (healthy macroscopic brain tissue), without contact to a tumor margin, derived from necessary corticotomy in the approach pathway in cases of subcortical located lesions.
Specimen samples from the tumor margins were not obtained from meningiomas, schwannomas, or other extra-axial lesions, but only from malignant intra-axial tumors to test the CNN algorithm reported here.
Evaluation of CNN models and prediction probabilities
Three CNN models were trained and tested with different random seeds on the same dataset for analysis. Examination and evaluation demonstrated a tumor mean probability of 73.9 (± 33.2) by the first CNN, 76.9 (± 35.0) by the second CNN, and 76.5 (± 33.7) by the third CNN. The interrater analysis of the probability values between the three CNN showed excellent reliability with an ICC value of 0.962 (99% CI 0.953–0.969).
The mean probability for non-tumor was 18.9 (± 33.1) for the first CNN, 18.0 (± 32.7) for the second, and 18.2 (± 31.6) for the third. The ICC was found likewise excellent with a value of 0.977 (99% CI 0.973–0.981) between the three CNNs.
The mean probability for the low-quality output class was 7.2 (± 15.2) for the first, 5.1 (± 16.4) for the second, and 5.3 (± 15.7) for the third CNN. The ICC for the low-quality output class between the CNNs showed excellent interrater reliability of 0.914 (99% CI 0.895–0.929) (further information see Table 2).
In the qualitative image inspection per viewing, some locations of the analyzed SRH images on patch size level showed some differences in the corresponding patch prediction output between the three CNNs but did not affect the overall prediction output of each SRH image. There were no statistically significant differences found between the three CNNs probability predictions regarding the evaluation of tumor tissue (p = 0.231), non-tumor (p = 0.052), and low quality (p = 0.423) SRH images (see Table 1; Fig. 3).
Diagnosis classification by the CNN models and neuropathologist
The neuropathologist established it as the ground truth, considering the majority area of the SRH image for the corresponding annotation, thus identifying 462 (78.0%) SRH images as tumor, 113 (19.1%) as non-tumor (including 34 SRH image negative controls from the surgical access, and tumor negative SRH images derived from tumor margins), and 17 (2.9%) as of low quality.
If one of the three diagnostic classes contained the highest probability prediction value, this class was chosen as the final diagnosis. After summation, the first CNN identified 459 (77.5%) SRH images as a tumor, 115 (19.4%) as non-tumor, and 18 (3%) as low quality.
The second CNN evaluated 456 (77%) SRH images as a tumor, 117 (19.8%) as non-tumor, and 19 (3.2%) as low quality.
Diagnostic quality of CNN models
Due to the sensitivity and specificity of ROC analysis, a comparable diagnostic quality could be reached in reference to the independent neuropathological evaluation. In detail, the ROC-AUC of the first CNN was demonstrated as excellent with 0.888 (99% CI 0.838–0.938) for tumor, 0.862 (99% CI 0.789–0.934) for non-tumor, and 0.837 (99% CI 0.576–1.097) for low-quality SRH images.
The ROC-AUC in the second CNN was excellent with 0.895 (99% CI 0.843–0.948) for tumor, 0.879 (99% CI 0.810–0.947) for non-tumor, and outstanding with 0.902 (99% CI 0.751–1.053) for low-quality SRH images.
The ROC-AUC of the third CNN was excellent with 0.882 (99% CI 0.826–0.938) for tumor, 0.876 (99% CI 0.810–0.941) for non-tumor, and 0.878 (99% CI 0.707–1.050) for low-quality SRH images (see Fig. 4).
Agreement between CNN models and neuropathologist
The lowest inter-rater agreement was found between the first CNN and neuropathological assessment in 86.5% SRH images with a moderate Cohen's Kappa (κ = 0.572).
Compared to the neuropathological evaluation, the second CNN showed an inter-rater consistency in 87.3% SRH images with a substantial Cohen’s Kappa (κ = 0.607).
The best Inter-rater agreement was found between the neuropathological evaluation and the evaluation of the third retrained CNN with 89.6% out of 402 SRH images with correct prediction and differentiation of tumor and non-tumor tissue. There was a substantial Cohen's Kappa (κ = 0.671) (see Table 1).
Internal consistency analysis
After analysis of the inter-rater agreement between CNNs and neuropathologist, the internal consistency of the three random area SRH images within the same slide (n = 132) was investigated.
The third CNN, which presented the strongest agreement compared to the neuropathological evaluation, demonstrated an excellent IC in 90.2% (Cα = 0.942) of the specimen slides. Here, the qualitative SRH image analysis with a review of the overlaying heatmap of the CNN showed minor differences in some areas on patch size level, which did not affect the overall prediction output (Further information see Fig. 5 as Additional file 1: Tables S1 and S2).
One of the main challenges in neurosurgery remains to distinguish tumorous tissue safely and efficiently from non-tumorous brain tissue to enable successful stereotactic biopsies and surgically complete resections. In addition, reliable quick histopathological diagnostics are increasingly needed for intraoperative decision-making and gain importance, e.g. for applying local adjuvant treatment options such as chemotherapeutic wafers  or intraoperative radiotherapy [12, 13]. The use of conventional rapid frozen section diagnostics with tissue processing and sectioning and white light microscopy by the (neuro)pathologist is the gold standard of care but involves a physical and time-consuming disruption of the surgical workflow . Despite efficiency optimization of the process over years it remains user-dependent, costly- and time-consuming with 20 to 30 min turnaround time, in some cancer centers even longer, and needs high resource requirements due to interdepartmental logistical challenges [3, 14]. Furthermore, destructive, freezing, and compressing artifacts can occur during tissue processing . Lastly, a (neuro)pathology unit is not always available in every surgical tumor center.
As an alternative histopathological assessment pathway the digital SRS microscopy has recently been introduced in the clinical setting as a label- and dye-free microscopic chemical imaging technique for unprocessed tissue specimens [8, 9]. The additive and encouraging use of AI applications with SRH microscopy has already been reported, allowing rapid prediction of tumor entity and glial dignity with encouraging agreement with the neuropathologist's diagnosis [4, 5]. Hollon et al. recently demonstrated an SRH-based CNN to distinguish between predefined tumor entities and dignity  but was limited in a qualitative conclusion on whether tumor or no tumor tissue is present. A qualitative assessment of a CNN that distinguishes between tumor and non-tumor has not yet been reported.
We developed three CNNs trained and tested on the same dataset and examined their diagnostic performance, reliability between CNNs, between CNNs and a neuropathologist, and the IC of multiple random areas within a sample.
Using cellular and nuclear morphological features, our trained and deployed CNNs were able to accurately identify and discriminate tumor from non-tumorous brain tissue or judge them as low-quality tissue by analyzing SRH images in a fully automated pipeline.
No significant differences in terms of probability prediction values of the tumor, non-tumor, and low-quality class were found between the three trained CNNs. In this context, previous studies show similar values but did not report detailed analyses of prediction probability values of several CNN models [4, 5, 14, 15]. We were able to demonstrate excellent concordance of the diagnosis prediction between the three CNNs, although some differences in qualitative image inspection were found at the patch size level, which did not negatively affect the overall prediction output of the whole SRH image.
Furthermore, in the reliability analysis between the CNNs and the independent neuropathological evaluation as the gold standard, the third developed CNN demonstrated the highest inter-rater agreement with approximately 90% and showed excellent diagnostic sensitivity and specificity regarding the evaluation of tumor, non-tumor tissue, and low-quality SRH images. It was therefore chosen for further validation and analysis. Other studies showed similar overall accuracy as well as ROC-AUC ranging from approximately 84 to 97% between the CNN- and neuropathological-based SRH image analysis [4, 5, 15, 16]. However, it should be noted that the CNN deployed in these studies had different output classes with different loss functions.
The question has been previously raised whether multiple scans in different areas of the same sample may even increase the diagnostic sensitivity. However, the intra-sample consistency throughout multiple areas has not yet been addressed in previous studies. In that regard, we found an excellent consistency of the applied prediction output between the three different random areas from the corresponding specimen slide of tumor, tumor margin area, or healthy brain tissue. Even if small differences were observed on patch size level, the renormalized summation could anyway result in a consistent diagnosis. In the evaluation of IC, the third trained CNN with the highest agreement in accordance with the neuropathological evaluation also showed an excellent agreement between three random areas within a specimen sample in 90.2%. This suggests that multiple scans within one sample in random areas may be deemed no longer necessary, resulting in an additional intraoperative time-saving.
Despite intraoperative technological advances like neuronavigation, intraoperative MRI-guidance, fluorescence-guided surgery, ultrasound modalities, and recently augmented reality applications [17,18,19,20,21,22,23], the surgeon still relies on rapid intraoperative histopathological diagnoses on a qualitative microscopic level for decision-making, as the aforementioned techniques have diverse limitations such as being prone to potential brain shift, limited resolution sensitivity and signal specificity as well as dependent to the varying uptake of fluorescent dyes used [18, 24].
Alternative advanced techniques to our presented AI-based intraoperative optical laser system for qualitative diagnostics are the intraoperative confocal laser endomicroscopy. It provides noninvasive real-time imaging. However, this technique is limited by the need for potential toxic fluorophores for contrast-enhancement and microstructure visualization . Moreover, this technique requires intense user training and merely provides non-static gray scale confocal microscopic images [26,27,28], which is challenging imaging data to interpret even for trained neuropathologists. In comparison, the CNN-based label-free SRH image acquisition and analysis studied here provides user independently fast feedback, and applicability in the operating room ex-situ and digital images familiarly color-coded like conventional H&E staining. The immediate assessment and connectivity through digital image transfer systems may even provide rapid image and analysis review by the (neuro)pathologist remotely.
Further randomized prospective studies in comparison to conventional histopathological workflows are mandatory for more detailed analysis and validation to establish such a workflow as a potential standard of care on a sustainable basis.
A limitation of our study is that the neuropathological evaluation and, thus, assessment of the ground truth for CNN comparison was performed by only one board-certified neuropathologist. However, all pathological reviews of the SRH images have been done by the same trained independent neuropathology attending with over 20 years of experience minimizing the systematic error. In addition, the diagnosis strategies differed between the CNNs (highest probability prediction value) and the neuropathologist (binary value), which could influence the final comparison and therefore the accuracy of comparison between both. Since the CNN algorithm sums the patches and averages their prediction values for each of the three output classes, rather than providing the highest tumor prediction value of any of the patches, the sensitivity of CNN-based tumor probability diagnosis is, thus, limited in images containing very small tumor infiltrations. That is, in particular, true in cases when single or very few patches predict tumor—equivalent to tumor infiltration in the low micrometer range—while the rest of the SRH image is free of tumor. However, the averaging approach minimizes the misdiagnosis of the sample due to potential artifactual tumor detections in individual patches given the small size of a single patch (14.2 × 13.8 µm) relative to the whole SRH image (1.7 × 1.8 mm). In order to still enable the image reader’s correct interpretation of the CNN results in those cases, SRH image overlaying probability maps are provided in addition and spatially color-code these small tumor areas for a focused look where the CNN has detected tumor.
Additionally, it reports on a large dataset representing the entire spectrum of neurooncological diagnoses.
In conclusion, CNN-based intraoperative evaluation of SRH images enables label-free tumor detection, irrespective of dignity and entity, in a fully automated pipeline and provides a reproducible alternative pathway for intraoperative decision-making in neurooncological surgery, independent of the conventional histopathologic frozen section examination.
Availability of data and materials
The datasets used and/or analyzed during the current study are available from V.N. on reasonable request.
Convolutional neural network
Intraclass correlation coefficient
Machine learning algorithm
Stimulated Raman scattering
Stimulated Raman histology
Receiver operating characteristic
Area under the ROC curve
Sullivan R, Alatise OI, Anderson BO, Audisio R, Autier P, Aggarwal A et al (2015) Global cancer surgery: delivering safe, affordable, and timely cancer surgery. Lancet Oncol 16:1193–1224
Laakman JM, Chen SJ, Lake KS, Blau JL, Rajan DA, Samuelson MI et al (2021) Frozen section quality assurance. Am J Clin Pathol 156:461–470
Novis DA, Zarbo RJ (1997) Interinstitutional comparison of frozen section turnaround time. A College of American Pathologists Q-Probes study of 32868 frozen sections in 700 hospitals. Arch Pathol Lab Med 121:559–67
Hollon TC, Pandian B, Urias E, Save AV, Adapa AR, Srinivasan S et al (2020) Rapid, label-free detection of diffuse glioma recurrence using intraoperative stimulated Raman histology and deep neural networks. Neuro Oncol 23:144–155
Hollon TC, Pandian B, Adapa AR, Urias E, Save AV, Khalsa SSS et al (2020) Near real-time intraoperative brain tumor diagnosis using stimulated Raman histology and deep neural networks. Nat Med 26:52–58
Hollon TC, Lewis S, Pandian B, Niknafs YS, Garrard MR, Garton H et al (2018) Rapid intraoperative diagnosis of pediatric brain tumors using stimulated raman histology. Cancer Res 78:278–289
Hollon T, Orringer DA (2021) Label-free brain tumor imaging using Raman-based methods. J Neurooncol 151:393–402
Orringer DA, Pandian B, Niknafs YS, Hollon TC, Boyle J, Lewis S et al (2017) Rapid intraoperative histology of unprocessed surgical specimens via fibre-laser-based stimulated Raman scattering microscopy. Nat Biomed Eng 1:0027
Freudiger CW, Min W, Saar BG, Lu S, Holtom GR, He C et al (2008) Label-free biomedical imaging with high sensitivity by stimulated Raman scattering microscopy. Science 322:1857–1861
Hosmer DW, Lemeshow S, Sturdivant RX. Applied logistic regression; 2013.
Chowdhary SA, Ryken T, Newton HB (2015) Survival outcomes and safety of carmustine wafers in the treatment of high-grade gliomas: a meta-analysis. J Neurooncol 122:367–382
Giordano FA, Wenz F, Petrecca K (2016) Rationale for intraoperative radiotherapy in glioblastoma. J Neurosurg Sci 60:350–356
Sethi A, Emami B, Small W, Thomas TO (2018) Intraoperative radiotherapy with INTRABEAM: technical and dosimetric considerations. Front Oncol 6:66. https://doi.org/10.3389/fonc.2018.00074
St John ER, Al-Khudairi R, Ashrafian H, Athanasiou T, Takats Z, Hadjiminas DJ et al (2017) Diagnostic accuracy of intraoperative techniques for margin assessment in breast cancer surgery: a meta-analysis. Ann Surg 265:300–310
Hollon TC, Orringer DA (2020) An automated tissue-to-diagnosis pipeline using intraoperative stimulated Raman histology and deep learning. Mol Cell Oncol 7:1736742
Jiang C, Bhattacharya A, Linzey JR, Joshi RS, Cha SJ, Srinivasan S et al (2022) Rapid automated analysis of skull base tumor specimens using intraoperative optical imaging and artificial intelligence. Neurosurgery. https://doi.org/10.1227/neu.0000000000001929
Rogers CM, Jones PS, Weinberg JS (2021) Intraoperative MRI for brain tumors. J Neurooncol 151:479–490
Gerard IJ, Kersten-Oertel M, Hall JA, Sirhan D, Collins DL (2020) Brain shift in neuronavigation of brain tumors: an updated review of intra-operative ultrasound applications. Front Oncol 10:618837
Inoue D, Cho B, Mori M, Kikkawa Y, Amano T, Nakamizo A et al (2013) Preliminary study on the clinical application of augmented reality neuronavigation. J Neurol Surg A Cent Eur Neurosurg 74:71–76
Stummer W, Pichlmeier U, Meinel T, Wiestler OD, Zanella F, Reulen H-J et al (2006) Fluorescence-guided surgery with 5-aminolevulinic acid for resection of malignant glioma: a randomised controlled multicentre phase III trial. Lancet Oncol 7:392–401
Unsgaard G, Ommedal S, Muller T, Gronningsaeter A, Nagelhus Hernes TA. Neuronavigation by intraoperative three-dimensional ultrasound: initial experience during brain tumor resection. Neurosurgery. 2002;50:804–12; discussion 812.
Mahboob S, McPhillips R, Qiu Z, Jiang Y, Meggs C, Schiavone G et al (2016) Intraoperative ultrasound-guided resection of gliomas: a meta-analysis and review of the literature. World Neurosurg 92:255–263
Noh T, Mustroph M, Golby AJ (2021) Intraoperative imaging for high-grade glioma surgery. Neurosurg Clin N Am 32:47–54
Drug Approval Package: Gleolan (aminolevulinic acid hydrochloride) [Internet]. [cited 2022 Apr 19]. https://www.accessdata.fda.gov/drugsatfda_docs/nda/2017/208630Orig1s000TOC.cfm
Restelli F, Pollo B, Vetrano IG, Cabras S, Broggi M, Schiariti M et al (2021) Confocal laser microscopy in neurosurgery: state of the art of actual clinical applications. J Clin Med 10:2035
Belykh E, Zhao X, Ngo B, Farhadi D, Byvaltsev V, Eschbacher J et al (2020) Intraoperative confocal laser endomicroscopy ex vivo examination of tissue microstructure during fluorescence-guided brain tumor surgery. Front Oncol 10:66
Eschbacher J, Martirosyan NL, Nakaji P, Sanai N, Preul MC, Smith KA et al (2012) In vivo intraoperative confocal microscopy for real-time histopathological imaging of brain tumors. J Neurosurg 116:854–860
Keerl R, Weber RK, Draf W, Wienke A, Schaefer SD (2004) Use of sodium fluorescein solution for detection of cerebrospinal fluid fistulas: an analysis of 420 administrations and reported complications in Europe and the United States. Laryngoscope 114:266–272
Open Access funding enabled and organized by Projekt DEAL. We acknowledge open access publishing funding support from the DFG (German Research Foundation, grant no. 491454339). The research received no other external funding.
Ethics approval and consent to participate
Ethical review and approval were given by the local ethics committee (Nr. 21-1238). Patient consent was given prior to inclusion in the study.
C.F., A.I.M., and F.K. are employees of Invenio Imaging Inc, the manufacturer of the SRH imaging system. C.F., A.I.M., and F.K. did not take part in any data analysis or influenced the report or interpretation of the results. No other conflicts of interest to declare.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
: Internal consistency analysis and overall agreement of the three residual Convolutional Neural Networks (CNNs). Table S2: Comparison of prediction probabilities of the three residual Convolutional Neural Networks (CNN) of the corresponding random areas of the same tumor samples in mean and standard deviation, and Intraclass correlation coefficient (ICC). Table S3: Subdivision of predictions of neuropathological evaluation (NP), residual convolutional neural network (CNN), and each interrater agreement (CP) based on the different entities.
About this article
Cite this article
Reinecke, D., von Spreckelsen, N., Mawrin, C. et al. Novel rapid intraoperative qualitative tumor detection by a residual convolutional neural network using label-free stimulated Raman scattering microscopy. acta neuropathol commun 10, 109 (2022). https://doi.org/10.1186/s40478-022-01411-x
- Brain tumor
- Tissue detection
- Stimulated Raman histology
- Deep learning
- Artificial intelligence