AI-Powered Fast Screening for Neoplastic Diseases in the Chilean Population

The COVID-19 pandemic has further emphasized the importance of telemedicine strategies. The growing demand for these services highlights the necessity to develop automated systems capable of categorizing consultations and prioritizing urgent cases. In Chile, however, some dermatological consultations are submitted with incomplete clinical information, and the high volume of cases often results in image quality issues. This study aims to support the screening of dermatological cases by evaluating completeness indicators through the detection of dermatological diseases in real-world clinical images. Additionally, it ensures quality by employing detection-related indicators, particularly the crop ratio, which quantifies the proportion of the image occupied by the detected lesion. We used object detection algorithms, specifically YOLOv8 architectures, to identify skin pathologies representative of the Chilean population. The data consisted of a large number of real telemedicine cases from Chile, ensuring a dataset representative of the Chilean population and potentially reflects the characteristics of the broader Latin American context. The results demonstrate that the models can accurately detect 82% of skin lesions in real-world clinical data. Expert validation indicates that 87% of the detected lesions are clinically relevant (95% CI: [83.7%, 90.3%]). Furthermore, our quality assurance measures using the crop ratio and expert validation reveal that images where the detected area constitutes 10% or less of the total image are considered inadequate for clinical evaluation. Our approach sets new precedents compared to previous deep learning models, which have predominantly relied on public datasets composed mainly of images from Caucasian populations, typically cleaned of noise and reviewed by experts, and therefore not fully representative of the variability and challenges encountered in real-world clinical screening contexts.