Wavenumber selection based analysis in Raman spectroscopy improves skin cancer diagnostic specificity

Improving skin cancer detection by Raman spectroscopy using convolutional neural networks and data augmentation

Background

Our previous studies have demonstrated that Raman spectroscopy could be used for skin cancer detection with good sensitivity and specificity. The objective of this study is to determine if skin cancer detection can be further improved by combining deep neural networks and Raman spectroscopy.

Patients and methods

Raman spectra of 731 skin lesions were included in this study, containing 340 cancerous and precancerous lesions (melanoma, basal cell carcinoma, squamous cell carcinoma and actinic keratosis) and 391 benign lesions (melanocytic nevus and seborrheic keratosis). One-dimensional convolutional neural networks (1D-CNN) were developed for Raman spectral classification. The stratified samples were divided randomly into training (70%), validation (10%) and test set (20%), and were repeated 56 times using parallel computing. Different data augmentation strategies were implemented for the training dataset, including added random noise, spectral shift, spectral combination and artificially synthesized Raman spectra using one-dimensional generative adversarial networks (1D-GAN). The area under the receiver operating characteristic curve (ROC AUC) was used as a measure of the diagnostic performance. Conventional machine learning approaches, including partial least squares for discriminant analysis (PLS-DA), principal component and linear discriminant analysis (PC-LDA), support vector machine (SVM), and logistic regression (LR) were evaluated for comparison with the same data splitting scheme as the 1D-CNN.

Results

The ROC AUC of the test dataset based on the original training spectra were 0.886±0.022 (1D-CNN), 0.870±0.028 (PLS-DA), 0.875±0.033 (PC-LDA), 0.864±0.027 (SVM), and 0.525±0.045 (LR), which were improved to 0.909±0.021 (1D-CNN), 0.899±0.022 (PLS-DA), 0.895±0.022 (PC-LDA), 0.901±0.020 (SVM), and 0.897±0.021 (LR) respectively after augmentation of the training dataset (p<0.0001, Wilcoxon test). Paired analyses of 1D-CNN with conventional machine learning approaches showed that 1D-CNN had a 1-3% improvement (p<0.001, Wilcoxon test).

Conclusions

Data augmentation not only improved the performance of both deep neural networks and conventional machine learning techniques by 2-4%, but also improved the performance of the models on spectra with higher noise or spectral shifting. Convolutional neural networks slightly outperformed conventional machine learning approaches for skin cancer detection by Raman spectroscopy.

Multimodal Raman spectroscopy and optical coherence tomography for biomedical analysis

Optical techniques hold great potential to detect and monitor disease states as they are a fast, non-invasive toolkit. Raman spectroscopy (RS) in particular is a powerful label-free method capable of quantifying the biomolecular content of tissues. Still, spontaneous Raman scattering lacks information about tissue morphology due to its inability to rapidly assess a large field of view. Optical Coherence Tomography (OCT) is an interferometric optical method capable of fast, depth-resolved imaging of tissue morphology, but lacks detailed molecular contrast. In many cases, pairing label-free techniques into multimodal systems allows for a more diverse field of applications. Integrating RS and OCT into a single instrument allows for both structural imaging and biochemical interrogation of tissues and therefore offers a more comprehensive means for clinical diagnosis. This review summarizes the efforts made to date toward combining spontaneous RS-OCT instrumentation for biomedical analysis, including insights into primary design considerations and data interpretation.

Classification of skin cancer using convolutional neural networks analysis of Raman spectra

BACKGROUND AND OBJECTIVE

Skin cancer is the most common malignancy in whites accounting for about one third of all cancers diagnosed per year. Portable Raman spectroscopy setups for skin cancer "optical biopsy" are utilized to detect tumors based on their spectral features caused by the comparative presence of different chemical components. However, low signal-to-noise ratio in such systems may prevent accurate tumors classification. Thus, there is a challenge to develop methods for efficient skin tumors classification.

METHODS

We compare the performance of convolutional neural networks and the projection on latent structures with discriminant analysis for discriminating skin cancer using the analysis of Raman spectra with a high autofluorescence background stimulated by a 785 nm laser. We have registered the spectra of 617 cases of skin neoplasms (615 patients, 70 melanomas, 122 basal cell carcinomas, 12 squamous cell carcinomas and 413 benign tumors) in vivo with a portable Raman setup and created classification models both for convolutional neural networks and projection on latent structures approaches. To check the classification models stability, a 10-fold cross-validation was performed for all created models. To avoid models overfitting, the data was divided into a training set (80% of spectral dataset) and a test set (20% of spectral dataset).

RESULTS

The results for different classification tasks demonstrate that the convolutional neural networks significantly (p<0.01) outperforms the projection on latent structures. For the convolutional neural networks implementation we obtained ROC AUCs of 0.96 (0.94 - 0.97; 95% CI), 0.90 (0.85-0.94; 95% CI), and 0.92 (0.87 - 0.97; 95% CI) for classifying a) malignant vs benign tumors, b) melanomas vs pigmented tumors and c) melanomas vs seborrheic keratosis respectively.

CONCLUSIONS

The performance of the convolutional neural networks classification of skin tumors based on Raman spectra analysis is higher or comparable to the accuracy provided by trained dermatologists. The increased accuracy with the convolutional neural networks implementation is due to a more precise accounting of low intensity Raman bands in the intense autofluorescence background. The achieved high performance of skin tumors classifications with convolutional neural networks analysis opens a possibility for wide implementation of Raman setups in clinical setting.

Improving SERS bioimaging of subcutaneous phantom in vivo with optical clearing

Surface-enhanced Raman scattering (SERS) has proven to be a promising technique for different types of imaging including preoperative and intraoperative in vivo tumor visualization. However, the strong scattering of the turbid tissue limits its use in subcutaneous areas. In this article, we used an optical clearing technique to improve the SERS signal from a subcutaneous tumor phantom. The phantom is a 2 mm sphere of calcium alginate with incorporated petal-like gap-enhanced Raman tags. The use of optical clearing increases the SERS signal target-to-background ratio for 5 times and allow to decrease the total imaging time for at least 10 times. In addition, SERS imaging assisted with optical clearing made it possible to more precisely determine the shape and boundaries of the implanted phantom. The combination of optical clearing and SERS is a promising strategy for the clinical imaging of subcutaneous objects that are usually shielded by dermal tissue.

Emerging Minimally Invasive Technologies for the Detection of Skin Cancer

With the increasing incidence of skin cancer, many noninvasive technologies to detect its presence have been developed. This review focuses on reflectance confocal microscopy (RCM), optical coherence tomography (OCT), high-frequency ultrasound (HFUS), electrical impedance spectroscopy (EIS), pigmented lesion assay (PLA), and Raman spectroscopy (RS) and discusses the basic principle, clinical applications, advantages, and disadvantages of each technology. RCM provides high cellular resolution and has high sensitivity and specificity for the diagnosis of skin cancer. OCT provides lower resolution than RCM, although its evaluable depth is deeper than that of RCM. RCM and OCT may be useful in reducing the number of unnecessary biopsies, evaluating the tumor margin, and monitoring treatment response. HFUS can be mainly used to delineate tumor depths or margins and monitor the treatment response. EIS provides high sensitivity but low specificity for the diagnosis of skin malignancies. PLA, which is based on the genetic information of lesions, is applicable for the detection of melanoma with high sensitivity and moderate-to-high specificity. RS showed high accuracy for the diagnosis of skin cancer, although more clinical studies are required. Advances in these technologies for the diagnosis of skin cancer can lead to the realization of optimized and individualized treatments.

Blood plasma resonance Raman spectroscopy combined with multivariate analysis for esophageal cancer detection

We herein report a novel, reliable and inexpensive method for detecting esophageal cancer using blood plasma resonance Raman spectroscopy combined with multivariate analysis methods. The blood plasma samples were divided into late stage cancer group (n = 164), early stage cancer group (n = 35) and normal group (n = 135) based on clinical pathological diagnosis. Using a specially designed quartz capillary tube as sample holder, we obtained higher quality resonance Raman spectra of blood plasma than existing method. The study demonstrated that the carotenoids levels in blood plasma were reduced in esophageal cancer patients. The area under the receiver operating characteristic curve (and 95% confidence interval) calculated by wavenumber selection and principal component analysis combined with linear discriminant analysis (PC-LDA) algorithm were 0.894 (0.858-0.929), 0.901 (0.841-0.960) and 0.871 (0.799-0.942) for differentiating late cancer from normal, late cancer from early cancer, and early cancer from normal respectively. The contribution from the two carotenoids wavenumber regions of 1155 and 1515 cm^-1 were more than 84.2%. The results show that the plasma carotenoids could be a potential biomarker for screening esophageal cancer using resonance Raman spectroscopy combined with wavenumber selection and PC-LDA algorithms.

In depth investigation of collagen non-enzymatic glycation by Raman spectroscopy

Non-enzymatic glycation is a post-translational modification of long-lived matrix proteins such as type I collagen. It occurs during aging and leads to the formation of advanced glycation end-products (AGEs). AGE accumulation is associated with severe complications in chronic and age-related diseases. The assessment of modifications induced by this (patho)physiological process represents an interest in biology and medicine for a better patient care. The objective of our work was to position the interest of Raman spectroscopy in the quantification of collagen glycation. Two types of in vitro glycation were used by incubating collagen samples, at different durations, with ribose or glyoxylic acid; these reducing agents acting on the chemical specificity of the glycation reaction. Glycation efficiency was evaluated by the liquid chromatography coupled to tandem mass spectrometry (LC-MS/MS) quantification of carboxymethyllysine (CML) and pentosidine, which are among the most studied AGEs. Raman data were processed by PCA coupled to validity indices and Lasso regression as multivariate analysis tools. Regression models were constructed by considering the LC-MS/MS results as reference values. A marked variability was observed within the Raman datasets making difficult the identification of spectral differences between control and ribose-treated collagen samples. By taking advantage of the chemical specificity of the glyoxylic acid treatment leading to CML formation, on one hand, and the feature selection included in the Lasso algorithm, on the other hand, Raman markers associated with glycation were identified. The assigned vibrations corresponded to modifications of side chains of collagen. In addition, a threshold of CML concentration was determined as quantitative indicator of the applicability of Raman spectroscopy for potential patient follow-up purposes. Although lacking in sensitivity to directly detect AGEs in collagen, Raman spectroscopy allows to highlight the molecular modifications of collagen induced by glycation.

In vivo diagnosis of skin cancer with a portable Raman spectroscopic device

In this study, we performed in vivo diagnosis of skin cancer based on implementation of a portable low-cost spectroscopy setup combining analysis of Raman and autofluorescence spectra in the near-infrared region (800-915 nm). We studied 617 cases of skin neoplasms (615 patients, 70 melanomas, 122 basal cell carcinomas, 12 squamous cell carcinomas and 413 benign tumors) in vivo with a portable setup. The studies considered the patients examined by GPs in local clinics and directed to a specialized Oncology Dispensary with suspected skin cancer. Each sample was histologically examined after excisional biopsy. The spectra were classified with a projection on latent structures and discriminant analysis. To check the classification models stability, a 10-fold cross-validation was performed. We obtained ROC AUCs of 0.75 (0.71-0.79; 95% CI), 0.69 (0.63-0.76; 95% CI) and 0.81 (0.74-0.87; 95% CI) for classification of a) malignant and benign tumors, b) melanomas and pigmented tumors and c) melanomas and seborrhoeic keratosis, respectively. The positive and negative predictive values ranged from 20% to 52% and from 73% to 99%, respectively. The biopsy ratio varied from 0.92:1 to 4.08:1 (at sensitivity levels from 90% to 99%). The accuracy of automatic analysis with the proposed system is higher than the accuracy of GPs and trainees, and is comparable or less to the accuracy of trained dermatologists. The proposed approach may be combined with other optical techniques of skin lesion analysis, such as dermoscopy- and spectroscopy-based computer-assisted diagnosis systems to increase accuracy of neoplasms classification.

Raman Microspectroscopic Investigation and Classification of Breast Cancer Pathological Characteristics

Breast cancer is one of the major cancers of women in the world. Despite significant progress in its treatment, an early diagnosis can effectively reduce its incidence rate and mortality. To improve the reliability of Raman-based tumor detection and analysis methods, we conducted an ex vivo study to unveil the compositional features of healthy control (HC), solid papillary carcinoma (SPC), mucinous carcinoma (MC), ductal carcinoma in situ (DCIS), and invasive ductal carcinoma (IDC) tissue samples. Following the identification of biological variations occurring as a result of cancer invasion, principal component analysis followed by linear discriminate analysis (PCA-LDA) algorithm were adopted to distinguish spectral variations among different breast tissue groups. The achieved results confirmed that after training, the constructed classification model combined with the leave-one-out cross-validation (LOOCV) method was able to distinguish the different breast tissue types with 100% overall accuracy. The present study demonstrates that Raman spectroscopy combined with multivariate analysis technology has considerable potential for improving the efficiency and performance of breast cancer diagnosis.

Feature Selection and Rapid Characterization of Bloodstains on Different Substrates

Establishing the precise timeline of a crime can be challenging as current analytical techniques used suffer from many limitations and are destructive to the body fluids encountered at crime scenes. Raman spectroscopy has demonstrated excellent potential in forensic science as it provides direct information about the structural and molecular changes without the need for processing or extracting samples. However, its current applicability is limited to pure body fluids, as signals from the substrate underlying these fluids greatly influence the current models used for age estimation. In this study, we utilized Raman spectroscopy to identify selective spectral markers that delineate the bloodstain age in the presence of interfering signals from the substrate. The pure bloodstains and the bloodstains on the substrate were aged for two weeks at 21 ± 2 ℃ in the dark. Least absolute shrinkage and selection operator (LASSO) regression was employed to guide the feature selection in the presence of interference from substrates to accurately predict the bloodstain age. Substrate-specific regression models guided by an automated feature selection algorithm yielded low values of predictive root mean square error (0.207, 0.204, 0.222 h in logarithmic scale) and high R² (0.924, 0.926, 0.913) on test data consisting of blood spectra on floor tile, facial tissue, and linoleum-polymer substrates, respectively. This framework for an automated feature selection algorithm relies entirely on pure bloodstain spectra to train substrate-specific models for estimating the age of composite (blood on substrate) spectra. The model can thus be easily applied to any new composite spectra and is highly scalable to new environments. This study demonstrates that Raman spectroscopy coupled with LASSO could serve as a reliable and nondestructive technique to determine the age of bloodstains on any surface while aiding forensic investigations in real-world scenarios.

Use of imaging techniques for melanocytic naevi and basal cell carcinoma in integrative analysis (Review)

Early detection of skin cancer is essential in order to obtain an improved prognosis. Clinicians need more objective and non-invasive examination methods to support their decision whether to biopsy or not tumoral lesions. These may include several imaging techniques such as dermoscopy, videodermoscopy, also known as sequential digital dermoscopy (SDD), computer-aided diagnosis (CAD), total body photography, imaging and high-frequency ultrasonography (HFUS), reflectance confocal microscopy, multiphoton tomography, electrical impedance spectroscopy, Raman spectroscopy, stepwise two-photon-laser spectroscopy and quantitative dynamic infrared. This review summarizes the current developments in the field of melanocytic lesions, such as naevi and basal cell carcinoma (BCC) imaging techniques. The aim was to collect and analyze data concerning types, indications, advantages and disadvantages of modern imaging techniques for in vivo skin tumor diagnosis. Two main methods were focused on, namely videodermoscopy and HFUS, which can be included in daily dermatologists' practice. In skin tumors HFUS allows the assessment of tumoral lesions with depth smaller than 1.5 cm, being described a correlation between ultrasonographic depth and the histologic index.

Raman Spectroscopy for Rapid Evaluation of Surgical Margins during Breast Cancer Lumpectomy

Failure to precisely distinguish malignant from healthy tissue has severe implications for breast cancer surgical outcomes. Clinical prognoses depend on precisely distinguishing healthy from malignant tissue during surgery. Laser Raman spectroscopy (LRS) has been previously shown to differentiate benign from malignant tissue in real time. However, the cost, assembly effort, and technical expertise needed for construction and implementation of the technique have prohibited widespread adoption. Recently, Raman spectrometers have been developed for non-medical uses and have become commercially available and affordable. Here we demonstrate that this current generation of Raman spectrometers can readily identify cancer in breast surgical specimens. We evaluated two commercially available, portable, near-infrared Raman systems operating at excitation wavelengths of either 785 nm or 1064 nm, collecting a total of 164 Raman spectra from cancerous, benign, and transitional regions of resected breast tissue from six patients undergoing mastectomy. The spectra were classified using standard multivariate statistical techniques. We identified a minimal set of spectral bands sufficient to reliably distinguish between healthy and malignant tissue using either the 1064 nm or 785 nm system. Our results indicate that current generation Raman spectrometers can be used as a rapid diagnostic technique distinguishing benign from malignant tissue during surgery.

Raman spectroscopy and chemometrics: A potential universal method for diagnosing cancer

Cancer is the second-leading cause of death worldwide. It affects an unfathomable number of people, with almost 16 million Americans currently living with it. While many cancers can be detected, current diagnostic efforts exhibit definite room for improvement. It is imperative that a person be diagnosed with cancer as early on in its progression as possible. An earlier diagnosis allows for the best treatment and intervention options available to be presented. Unfortunately, existing methods for diagnosing cancer can be expensive, invasive, inconclusive or inaccurate, and are not always made during initial stages of the disease. As such, there is a crucial unmet need to develop a singular universal method that is reliable, cost-effective, and non-invasive and can diagnose all forms of cancer early-on. Raman spectroscopy in combination with advanced statistical analysis is offered here as a potential solution for this need. This review covers recently published research in which Raman spectroscopy was used for the purpose of diagnosing cancer. The benefits and the risks of the methodology are presented; however, there is overwhelming evidence that suggests Raman spectroscopy is highly suitable for becoming the first universal method to be used for diagnosing cancer.

High-Throughput Cell Imaging and Classification by Narrowband and Low-Spectral-Resolution Raman Microscopy

We investigated the use of narrowband Raman spectra for rapid label-free molecular imaging aimed at cell classification using principal component regression and linear discriminant analysis. In the classification of breast nontumorigenic epithelial and cancer cell lines, the classification accuracies using a spectral range of 100 cm^-1 were equivalent to or better than that with using the fingerprint and high-wavenumber regions. Narrowing the Raman spectral range for analysis allows reduction of the charge-coupled device (CCD) pixels required for spectrum detection, resulting in the improvement of image acquisition speed with adequate classification accuracy. Our measurements revealed that the wavenumber region at 1397-1501 cm^-1 can provide molecular information sufficient for cell classification without causing notable errors in the baseline-correction. A spectral resolution of ∼9 cm^-1 was found to be sufficient to provide high accuracy in cell classification, which allowed us to apply pixel binning at the CCD readout for further acceleration of the imaging speed. As a result, the acquisition time for a 1200 × 1500 pixels Raman hyperspectral image at 1397-1501 cm^-1 was reduced to 21 min. Under this condition, different cell lines were classified at accuracies higher than 90%. The presented approach will improve throughput of cell and tissue analysis and classification using Raman spectroscopy and extend practical uses of Raman imaging in biology and medicine.

Confocal Raman microscopy combined with optical clearing for identification of inks in multicolored tattooed skin in vivo

Raman measurements applied on freshly tattooed porcine skin ex vivo showed a possibility of obtaining the ink pigment related information in the skin. Based on these results, confocal Raman microscopy was used to identify the tattoo ink pigments of different colors in multicolored tattooed human skin in vivo. The Raman signatures of tattoo ink pigments were unique. Therefore, it could be shown that the applied method is successful for the identification of the tattoo ink pigments in human skin in vivo down to depths of approx. 50 μm, which is sufficient to screen the entire epidermis and the top of the papillary dermis area on the forearm and leg skin sites. Additional application of the optical clearing technique in vivo by topical application of glycerol, combined with tape stripping removal of the uppermost stratum corneum layers and defatting allows the extension of depths of investigation in tattooed skin down to approx. 400 μm, i.e. to cover the entire papillary dermis and a large part of the reticular dermis. Thus, the tattoo ink pigments were identified in vivo and depth-dependently in human tattooed skin confirming their presence in the papillary and reticular dermis. The proposed non-invasive in vivo Raman screening combined with optical clearing for identifying the tattoo pigments in the dermis can be an important task preceding a laser-based tattoo removal procedure and for determining the optimal laser parameters.

Using Raman Spectroscopy to Detect and Diagnose Skin Cancer In Vivo

Raman spectroscopy provides a noninvasive bedside tool that captures unique optical signals via molecular vibrations in tissue samples. Raman theory was discovered at the beginning of the twentieth century, but it was not until the past few decades that it has been used to differentiate skin neoplasms. We provide a brief description of Raman spectroscopy for in vivo skin cancer diagnosis, including the physical principles underlying Raman spectroscopy, its advantages, typical spectra of skin pathologies, and its clinical application for aiding skin cancer diagnosis.

Influence of water content on Raman spectroscopy characterization of skin sample

We report that the Raman spectrum obtained from porcine skin varies significantly with the change of skin water content. At different water contents from 40 to 55 wt.%, the Raman spectra results using confocal Raman spectroscopy show that the spectral variation of porcine skin is highly affected by skin water content. Experimental data are consistent with the Monte Carlo calculation and it is proved that the intensity of the Raman spectrum depends on the angle distribution and collection efficiency of backscattered light from the sample surface for a varied water content. It is suggested that water content for a given skin sample should be controlled carefully to minimize errors and deviations in the Raman peak analyses.

Reflectance confocal microscopy of skin in vivo: From bench to bedside

Following more than two decades of effort, reflectance confocal microscopy (RCM) imaging of skin was granted codes for reimbursement by the US Centers for Medicare and Medicaid Services. Dermatologists in the USA have started billing and receiving reimbursement for the imaging procedure and for the reading and interpretation of images. RCM imaging combined with dermoscopic examination is guiding the triage of lesions into those that appear benign, which are being spared from biopsy, against those that appear suspicious, which are then biopsied. Thus far, a few thousand patients have been spared from biopsy of benign lesions. The journey of RCM imaging from bench to bedside is certainly a success story, but still much more work lies ahead toward wider dissemination, acceptance, and adoption. We present a brief review of RCM imaging and highlight key challenges and opportunities. The success of RCM imaging paves the way for other emerging optical technologies, as well-and our bet for the future is on multimodal approaches. Lasers Surg. Med. 49:7-19, 2017. © 2016 Wiley Periodicals, Inc.