Beyond conventional images: AI-driven biotechnologies for oral cancer diagnosis - a systematic review

Oral cancer (OC) represents a significant global health challenge, with traditional imaging techniques often falling short in early diagnosis. Recent advancements combining artificial intelligence (AI) and biotechnology have led to more accurate diagnostic outcomes. We conducted a systematic literature review of peer-reviewed original articles that utilized AI-driven biotechnology for early OC diagnosis. The studies were categorized into four groups: molecular biology, other biomarkers, spectral analysis, and multispectral autofluorescence lifetime imaging (maFLIM). We performed a comprehensive descriptive analysis at the group level and compared their diagnostic performances. Additionally, we employed an adapted Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) analysis to identify current limitations. A total of 42 studies were analyzed, yielding an overall accuracy of 87.9 % (range: 71.0 %-100.0 %). The molecular biology group exhibited the best performance, achieving an accuracy of 95.9 % (range: 95.0 %-96.7 %); spectroscopy also performed well, while maFLIM showed the poorest performance among the four groups. Quality assessments indicated significant risks in several domains. External validation was absent in 90.5 % of studies, and 20 % lacked a clearly defined model architecture. To validate model superiority, most studies compared against other state-of-the-art models, few studies compared different sample types, while others lacked comparisons. Additionally, there was a moderate to high risk in Patient Selection section due to unclear dataset composition, processing, and partitioning. Furthermore, traditional machine learning methods constituted 69.0 % of the studies, indicating a limited exploration of novel AI architectures. AI-driven biodiagnostic technologies demonstrate strong potential for the early diagnosis of OC. However, compared to other oncology fields, challenges such as limited sample sizes, insufficient validation and under-exploration of AI remain prevalent. Ongoing exploration of advanced deep learning techniques and multimodal approaches also holds promise for enhancing the clinical applicability of these technologies.