Oral cancer poses significant health challenges, with many patients diagnosed only at advanced stages. Recent advancements using convolutional neural networks (CNNs) promise to revolutionize the way oral cancers and precursors, such as oral epithelial dysplasia (OED), are detected. A groundbreaking study by researchers at the University of Melbourne and Optiscan Imaging Ltd has showcased how high-resolution confocal microscopy, paired with deep learning techniques, can facilitate real-time diagnoses of oral mucosal abnormalities.
Currently, the gold standard for detecting oral cancer involves invasive biopsy procedures, which are not only uncomfortable but also subject to classification errors based on human interpretation. Approximately two-thirds of patients are diagnosed too late to improve survival, with many facing extensive treatment options for advanced disease. This necessity for improved early detection is where CNNs come to the forefront.
The study enrolled fifty-nine patients with oral mucosal disorders between December 2020 and March 2022, utilizing confocal laser endomicroscopy to capture nearly 9,168 image frames. These images were processed using various pre-trained models of the CNN architecture known as
Inception-V3 to assess the quality of the images and classify them based on histopathological findings.
One of the standout performances of the CNNs recorded accuracy levels as high as 89.5% for image quality filtering alone. With contrasting agents like acriflavine and fluorescein, the models demonstrated exceptional ability to categorize lesions ranging from no dysplasia to high-grade dysplasia and OSCC. The fluorescein-specific model revealed impressive area under the curve (AUC) scores, ranging from 0.90 to 0.96, indicating its strong diagnostic performance across multiple classes. Conversely, the acriflavine model, though proficient at identifying certain lesion types, struggled with others, highlighting areas for improvement.
According to the study, "This study suggests tandem CNNs can provide highly accurate and rapid real-time diagnostic triage for in vivo assessment of high-risk oral mucosal disease.” Such rapid classification speeds, often cited at under 0.1 seconds per image, can significantly expedite the process of diagnosing oral diseases, allowing for immediate clinical feedback and reducing the reliance on traditional biopsy methods.
The findings also note the importance of identifying lesions inherently at risk of malignant transformation. With appropriate imaging techniques and CNN diagnostics, clinicians can track potentially malignant disorders more effectively, facilitating timely treatment interventions.
Previous methods of oral cancer detection using deep learning have largely focused on histopathology slides, which are far more invasive. This innovative approach combines the non-invasive capabilities of confocal microscopy with the power of AI to create potentially life-saving outcomes for patients.
The CNN models demonstrated excellence particularly at specific sites within the oral cavity, such as the floor of the mouth, which may possess unique structural features readily identifiable through advanced microscopy. This study paves the way for future research to explore the full diagnostic capabilities of CNNs, with the hope of enhancing the effectiveness and efficiency of oral cancer detection and management.
Summarizing this innovative approach, researchers assert, "Our findings indicate the models can significantly reduce the need for invasive biopsies by improving early detection and diagnostic accuracy.” The advent of AI and image processing techniques holds the promise of transforming preventative oncology, aiming to save lives through timely interventions.