Chondrogenic Cancer Grading by Combining Machine and Deep Learning with Raman Spectra of Histopathological Tissues

Gianmarco Lazzini, Mario D’Acunto

Raman spectroscopy (RS) is a promising tool for cancer diagnosis. In particular, in the last years several studies have demonstrated how the diagnostic performances of RS can be significantly improved by employing machine learning (ML) algorithms for the interpretation of Raman-based data. Recently, it has been demonstrated that RS can perform an accurate classification of chondrosarcoma tissues. Chondrosarcoma is a cancer of bones, that can occur in the soft tissues near the bones. It is normally characterized by three different malignant degrees and a benign counterpart, knows as enchondroma. In line with these findings, in this paper, we exploited ML algorithms to distinguish, as well as possible, between the three grades of chondrosarcoma and to distinguish between chondrosarcoma and enchondroma. We obtained a high level of accuracy of classification by analyzing a dataset composed of a relatively small number of Raman spectra, collected in a previous study by one of the authors of this paper. Such spectra were acquired from micrometric tissue sections with a confocal Raman microscope. We tested the classification performances of a support vector machine (SVM) and a random forest classifier (RFC), as representatives of ML algorithms, and two versions of the multi-layer perceptron (MLPC) as representatives of deep learning (DL). These models, especially RFC and MLPC, showed excellent classification performances, with accuracy reaching 99.7%. This outcome makes the aforementioned models a promising route for future improvements of diagnostic devices focused on detecting cancerous bone tissues. Alongside the diagnostic purpose, the aforementioned approach allowed us to identify characteristic molecules, i.e., amino acids, nucleic acids, and bioapatites, relevant for obtaining the final diagnostic response, through the use of a tool named by us Raman Band Identification (RBI). The method to evaluate RBI is the most important contribution of this paper, because RBI could represent a relevant parameter for the identification of biochemical processes on the basis of the tumor progression and associated malignant degree. In turn, the spectral bands highlighted by RBI could provide precious indicators in an attempt to restrict the spectral acquisition to specific Raman bands. This last objective could help to reduce the amount of experimental data needed to obtain an accurate final grading outcome, with a consequent reduction in the computational cost.