Termografia e comparação da performance de algoritmos de Machine Learn para análise preditiva de diagnóstico diferencial de queilite actínica e carcinoma de células escamosas de lábio

Abstract

Background: Non-invasive and non-ionizing infrared medical thermography is an auxiliary test used to detect abnormalities in the temperature profile of patients to discover an underlying pathology or inflammation. Thus, this method is of great value for the diagnostic aid of lip lesions. The lip is a region commonly attacked by ultraviolet radiation, especially in individuals exposed chronically to the sun without the proper use of sunscreens. Among the lesions that develop in the region of the lips, especially in the lower lip, are actinic cheilitis (AC) and lower lip squamous cell carcinoma (LLSCC), whose conclusive diagnosis is made through incisional biopsy. Objectives: To evaluate the use of thermography in the diagnosis of AC and LLSCC, through machine learning. Materials and Methods: Study was observational, with non probabilistic sampling technique, for convenience. The sample consisted of 65 individuals, grouped into: test group (33 individuals with lesions on the lip) and control group (32 individuals without lesions on the lip). Thermographic images were obtained using the FLIR T650sc® camera. Using a cooccurrence matrix with 9 attributes, the final sample corresponded to 312 thermographic images. Incisional biopsies were also performed for microscopic analysis of AC A and LLSCC cases, using the WHO morphological classification (2005; 2017). Python was used to develop the algorithm. For machine learning, three classifiers (KNN, SVM and DNN) were used. The Panda library was used for data treatment and analysis, and Scikit learn for machine learning. The confusion matrix was used to assess accuracy, sensitivity and precision. Results: A total of 27 AC cases (81.8%) with a strong predilection for light-skinned men (n = 18; 66.7% (n = 21; 77.8%) and mean age of 52.3 ± 13.9 were observed. For cases of CCELI, 6 (18.2%) were diagnosed, 66.6% (n = 4) of white individuals, all male, with an average age of 47.83 ± 14.66. These individuals, 63.6% (n = 21) worked chronically exposed to the sun, without the use of lip balm (n = 24; 71.7%) and 69.7% (n = 23) denied habits such as alcoholism and smoking presenting low educational level (n = 26; 78.8%). Concerning the morphological analysis, 48.2% (n = 13) cases of AC were diagnosed as hyperkeratosis and 37.0% (n = 10) as sligth epithelial dysplasia, and 83.3% (n = 5) as well-differentiated LLSCC from as recommended by WHO (2005). Among classifiers evaluated, DNN classifier showed major accuracy between test and control group (97.33%), and between AC and LLSCC (95.55%). DNN classifier exhibited only 6.4% false negatives to test group and none to control group. Between AC and LLSCC, DNN classifier exhibited non false negative to AC and only 2.5% false negatives to LLSCC. Conclusions: Through machine learning, using the RND classifier, it is suggested that thermography can be used as a health technology for the clinical diagnosis of QA and CCELI. In this way, numerous benefits, such as the early diagnosis of lip cancer and the reduction of public costs used for the treatment of this malignant neoplasm.