Artificial intelligence-based PRO score assessment in actinickeratoses from LC-OCT imaging using Convolutional NeuralNetworks

Background and Objectives:The histological PRO score (I-III) helps to assess themalignant potential of actinic keratoses (AK) by grading the dermal-epidermaljunction (DEJ) undulation. Line-field confocal optical coherence tomography (LC-OCT) provides non-invasive real-time PRO score quantification. From LC-OCTimaging data, training of an artificial intelligence (AI), using Convolutional Neu-ral Networks (CNNs) for automated PRO score quantification of AKin vivomay beachieved. Patients and Methods:CNNs were trained to segment LC-OCT images ofhealthy skin and AK. PRO score models were developed in accordance with thehistopathological gold standard and trained on a subset of 237 LC-OCT AK imagesand tested on 76 images, comparing AI-computed PRO score to the imagingexperts'visual consensus. Results:Significant agreement was found in 57/76 (75%) cases. AI-automatedgrading correlated best with the visual score for PRO II (84.8%) vs. PRO III (69.2%)vs. PRO I (66.6%). Misinterpretation occurred in 25% of the cases mostly due toshadowing of the DEJ and disruptive features such as hair follicles. Conclusions:The findings suggest that CNNs are helpful for automated PRO scorequantification in LC-OCT images. This may provide the clinician with a feasible toolfor PRO score assessment in the follow-up of AK.