Development of deep learning algorithms to discriminate giant cell tumors of bone from adjacent normal tissues by confocal Raman spectroscopy

Comparative study of machine-and deep-learning based classification algorithms for biomedical Raman spectroscopy (RS): case study of RS based pathogenic microbe identification

One key aspect pushing the frontiers of biomedical RS is dedicated machine- or deep- learning (ML or DL) algorithms. Yet, systematic comparative study between ML and DL algorithms has not been conducted for biomedical RS, largely due to the limited availability of open-source and large Raman spectra dataset. Therefore we compared typical ML partial least square-discriminant analysis (PLS-DA) and DL one dimensional convolution neural network (1D-CNN) based pathogenic microbe identification on 12,000 Raman spectra from six species of microbe (i.e., K. aerogenes (Klebsiella aerogenes), C. albicans (Candida albicans), C. glabrata (Candida glabrata), Group A Strep. (Group A Streptococcus), E. coli1 (Escherichia coli1), E. coli2 (Escherichia coli2)) when 100%, 75%, 50% and 25% of the 12,000 Raman spectra were retained. The total Raman dataset was analyzed with 80% split for training and 20% for testing. The 100% retained testing dataset accuracy, area under curve (AUC) of the receiver operating characteristic (ROC) curve were 95.25% and 0.997 for 1D-CNN, which are higher than those (89.42% and 0.979) of PLS-DA. Yet, PLS-DA outperforms 1D-CNN for 75%, 50% and 25% retained testing dataset. The resultant accuracies and AUCs demonstrated the performance reliance of PLS-DA and 1D-CNN on Raman spectra number. Besides, both loadings on the latent variables of PLS-DA and the saliency maps of 1D-CNN largely captured Raman peaks arising from DNA and proteins with comparable interpretability. The results of the current work indicated that both ML and DL algorithms should be explored for application-wise Raman spectra identification to select whichever with higher accuracies and AUCs.

Label-Free, Noninvasive Bone Cell Classification by Hyperspectral Confocal Raman Microscopy

Characterizing and identifying cells in multicellular in vitro models remain a substantial challenge. Here, we utilize hyperspectral confocal Raman microscopy and principal component analysis coupled with linear discriminant analysis to form a label-free, noninvasive approach for classifying bone cells and osteosarcoma cells. Through the development of a library of hyperspectral Raman images of the K7M2-wt osteosarcoma cell lines, 7F2 osteoblast cell lines, RAW 264.7 macrophage cell line, and osteoclasts induced from RAW 264.7 macrophages, we built a linear discriminant model capable of correctly identifying each of these cell types. The model was cross-validated using a k-fold cross validation scheme. The results show a minimum of 72% accuracy in predicting cell type. We also utilize the model to reconstruct the spectra of K7M2 and 7F2 to determine whether osteosarcoma cancer cells and normal osteoblasts have any prominent differences that can be captured by Raman. We find that the main differences between these two cell types are the prominence of the β-sheet protein secondary structure in K7M2 versus the α-helix protein secondary structure in 7F2. Additionally, differences in the CH2 deformation Raman feature highlight that the membrane lipid structure is different between these cells, which may affect the overall signaling and functional contrasts. Overall, we show that hyperspectral confocal Raman microscopy can serve as an effective tool for label-free, nondestructive cellular classification and that the spectral reconstructions can be used to gain deeper insight into the differences that drive different functional outcomes of different cells.

Multi-scale sequential feature selection for disease classification using Raman spectroscopy data

Raman spectroscopy (RS) optical technology promises non-destructive and fast application in medical disease diagnosis in a single step. However, achieving clinically relevant performance levels remains challenging due to the inability to search for significant Raman signals at different scales. Here we propose a multi-scale sequential feature selection method that can capture global sequential features and local peak features for disease classification using RS data. Specifically, we utilize the Long short-term memory network (LSTM) module to extract global sequential features in the Raman spectra, as it can capture long-term dependencies present in the Raman spectral sequences. Meanwhile, the attention mechanism is employed to select local peak features that were ignored before and are the key to distinguishing different diseases. Experimental results on three public and in-house datasets demonstrate the superiority of our model compared with state-of-the-art methods for RS classification. In particular, our model achieves an accuracy of 97.9 ± 0.2% on the COVID-19 dataset, 76.3 ± 0.4% on the H-IV dataset, and 96.8 ± 1.9% on the H-V dataset.

Euclidean distance-based Raman spectroscopy (EDRS) for the prognosis analysis of gastric cancer: A solution to tumor heterogeneity

The prognosis analysis of gastric cancer is critical for selection of treatments and development of advanced therapeutic methods. A prognosis approach that is accurate, fast, convenient, and of low cost for gastric cancers is in high demand. Raman spectroscopy is a label-free and non-destructive technique to provide molecular fingerprints of biological samples, holding promises for cancer prognosis. However, the major challenge of gastric cancer prognosis lies in the widely existing tumor heterogeneity, which leads to unexpected spectral variations within one type of samples. In this work, we have developed the Euclidean distance (ED)-based Raman spectroscopy (EDRS) method for the prognosis analysis of gastric cancer to eliminate the influence of tumor heterogeneity. Raman spectra were first collected on the slices of paraffin-preserved tumor tissues from gastric cancer patients. A standard spectrum to represent the 'worst prognostic tumor cells' was then established. The similarity between each spectrum of tissues and the standard spectrum was assessed by ED, to provide a direct assessment on the prognosis status. We have successfully classified the patients into poor and favorable prognosis groups, either based on the averaged regional ED values (sensitivity of 75 %, specificity of 96.8 %), or based on the minimal ED values at the patient level (sensitivity of 90 %, specificity of 100 %). EDRS was also investigated for survival analysis (AUC = 0.955), much better than the commonly applied post-neoadjuvant therapy (ypTNM) category (AUC = 0.718). Our work highlights EDRS as a rapid, accurate, low-cost and robust tool for heterogeneous cancer-related prognosis assessment and survival prediction, providing new insights for spectroscopic tumor analysis.