Dermatologist-level classification of malignant lip diseases using a deep convolutional neural network.

Area of Science:

• Artificial Intelligence in Dermatology
• Medical Image Analysis
• Oncology

Background:

• Deep convolutional neural networks (DCNNs) demonstrate dermatologist-level accuracy in classifying various skin diseases.
• Research is needed to assess DCNN performance for specific anatomical locations like the lips.
• Lip histology and morphology present unique characteristics compared to other skin areas.

Purpose of the Study:

• To evaluate the efficacy of a DCNN algorithm in classifying benign and malignant lip diseases.
• To compare the DCNN's diagnostic performance against dermatologists and non-specialist physicians.

Main Methods:

• A DCNN model (Inception-Resnet-V2) was trained on 1629 lip lesion images (743 malignant, 886 benign).
• Model performance was validated using independent datasets of 344 and 281 images from external hospitals.
• DCNN classifications were compared with diagnoses from 44 human participants, including board-certified dermatologists and medical students.

Main Results:

• The DCNN achieved an area under the curve (AUC) of 0.827 for the 344-image set and 0.774 for the 281-image set.
• Sensitivity and specificity for the DCNN were comparable to dermatologists in classifying malignancy.
• Non-dermatologists showed a significant improvement in diagnostic accuracy after consulting the DCNN's output (Youden index increased from 0.201 to 0.322).

Conclusions:

• DCNNs can accurately classify benign and malignant lip diseases, performing on par with expert dermatologists.
• The DCNN algorithm serves as a valuable tool for non-specialist physicians, enhancing their ability to differentiate lip conditions.
• This AI application holds significant clinical implications for improving diagnostic accuracy in rare lip diseases and supporting clinical decision-making.