Skip to main content
. 2023 Apr 5;13(7):1353. doi: 10.3390/diagnostics13071353

Table 3.

A list of articles on ML models for oral cancer diagnosis.

Ref. ML Approaches Used Data Set Computation Tools Features Extracted/Features Selected Feature Extraction Approach Key Contribution Limitations Performance Evaluation Metrics
[144] LR, DT, SVM, and K-NN 467 OSCC patients MATLAB R2020a Prognostic features PCA and bivariate analysis It will allow the clinicians to predict the progression of the disease. Genetic profiling, biomarker analysis, and sophisticated histopathology imaging were missing from this paper. Acc = 0.705, specificity = 0.841, sensitivity = 0.41
[163] SVM, GMM The 1194 cells were taken from 341 healthy and 429 OSF with dysplasia photos. Snake tool for image segmentation Hyperchromasia, and nuclear texture, 23 characteristics were derived from segmented biopsy pictures. Active contour method of gradient vector flow (GVF). A median filtering technique is suggested for image pre-processing to get rid of the noise. Expert’s topic expertise and the right image processing were absent. Acc = 99.66%
[164] CNN, Gabor filter, Random forests High-grade = 15 Low-grade = 25 and Healthy = 2 subjects. Computer aided automatic tools Texture-based features Gabor feature extraction The identification of keratin pearls and the segmentation of subepithelial and epithelial layers can be used for oral precancerous screening and OSCC grading, respectively. Very little research was carried out on cytopathological and histological pictures to identify the keratin pearl structure. Acc = 99.88%
[165] ANN 211 cases with OSCC were identified between 1990 and 2000. Statistical tools Age and gender of the patient during the time of diagnosis were considered when data were analyzed. Peri-tumoral inflammatory infiltrate with local recurrence This study’s goal was to ascertain whether patients with OSCC may have their 5-year survival rate and incidence rate of LR affected by the presence and grade of PTI It was not possible to determine involvement in other age-related cancers. Specificity = 90.59%, sensitivity = 67.74%, Acc = 78.56%
[166] LR, linear SVM A total of 34 patients were enlisted for tissue biopsies of suspected oral epithelial lesions. - Spectral, time-resolved, and autofluorescence features. Linear discriminant analysis They created a CAD system that used ML to automatically distinguish between malignant and healthy oral tissue using data from in vivo widefield maFLIM endoscopy. In this study, numerous spectra per individual were used as separate datasets, resulting in training and testing sets that were not genuinely independent. F1 score = 0.85, specificity = 74%, sensitivity = 94%
[131] SVM, RF, LR, and K-NN High-definition cytology photos Telectology platform Mitotic figures, hyperchromatic nucleus, multiple nuclei, etc. Field of view extraction method This study thus prove the value of tele cytology for accurate, remote diagnosis and the application of autonomous ANN-based assessment to increase its efficiency. According to the limitations of traditional cytology, OPML can only be recognized with a poor sensitivity of approximately 18%. It demonstrated an accuracy result of 84 to 86% in the identification of oral lesions
[167] LR, RF, SVM, NB 145 patients suffering from early stage OTSCC. GridSearchCV, StratifiedKFold, and sklearn Python tools. Simple clinical and pathologic characteristics linked to patients’ prognoses were the factors used for this investigation. - They proved that the best approach is not to create an application that blends ML algorithms with an EHR system. Lack of large training sets and samples. The best results were achieved by the random forest model (specificity = 75%: sensitivity = 85%; AUC = 0.786.
[168] KNN Using a cytology-on-a-chip method, 999 patients had OSCC and PMOLs. Data visualization tools, cytopathology tools 144 cellular/nuclear features were gathered from single-cell analyses. PCA The results of the present study demonstrated the benefit of a POC-amenable cytology platform that can detect and monitor oral lesions throughout the full spectrum of OED diagnoses. The present study was limited by the fact that past investigations of cytology adjuncts and POCOCT, in general, focused primarily on PMOL examination in secondary conditions or clinical settings, where malignant and dysplastic lesions could be significantly more prevalent compared to the primary clinical setting. Acc = 99.3 %