The facial iDScore project: artificial intelligence models for diagnosis support

Introduction: Artificial Intelligence (AI) is a set of techniques that are constantly being researched, improved and new ones invented. AI has two main subset of techniques that are machine learning and deep learning. Especially the second group of techniques is capable of analysing large amounts of very complex data such as images. AI originated around the 1950s and is now used on a daily basis in many different fields, including medicine. Diagnosing facial pigmented lesions is a challenging topic for dermatologists. Accurate diagnosis of these lesions is crucial for effective patient management, especially in dermatology where visual assessment plays a central role. Wrong diagnosis could bring to unnecessary excision or wrong treatment options. AI could play a key role in improving the diagnostic accuracy. Objective: This thesis aims to evaluate and compare the effectiveness of two distinct models – a traditional machine learning scoring model based on logistic regression and a more modern Convolutional Neural Network (CNN) model – in the diagnosis and management of facial pigmented lesions. Material and Methods: The study involved the collection and analysis of dermoscopic images of facial lesions excised for malignancy suspicious and histologically confirmed. Two malignant diagnoses, i.e. lentigo maligna and lentigo maligna melanoma, and five benign diagnoses, i.e. atypical nevi, pigmented actinic keratosis, solar lentigo, seborrheic keratosis, and seborrheic lichenoid keratosis were included. These images were provided by twelve European centres and then they were assessed by European dermatologists, and the data was used to develop and validate two diagnostic models. The first model is a score model based on logistic regression, incorporating dermoscopic patterns, lesion diameter, patient sex, and age. The second is a CNN model, leveraging deep learning techniques for image analysis. Both models were trained and tested in separated samples. CNN model was validated with a 5-fold cross validation. Results: 1197 images were collected and analysed. Patients mean age was 65.45±14.2, and the gender was equally distributed. About 40% of the database was constituted by malignant cases. 154 dermatologists evaluated the images, and the 50% of them had at least 4 years of experience in dermoscopy. The score model exhibited an Area Under the ROC (AUROC) of 83.5% in the training set and 79.4% in the testing set. In contrast, the CNN model showed a high sensitivity for melanoma diagnosis, with a consistent performance across various subsets. The CNN model's performance was comparatively higher in identifying melanoma but required a significantly larger dataset for training. The score model, while less technologically advanced, demonstrated practical applicability and reliable accuracy with a smaller dataset. Conclusion: Both models have advantages and limitations. The score model, grounded in traditional dermatological practice, benefits from its simplicity and direct applicability in clinical settings. On the other hand, the CNN model, despite its higher sensitivity in detecting melanoma, faces challenges in integration into clinical practice due to its need for extensive data and computational resources. In conclusion AI-based decision systems in the analysis of facial pigmented lesions could provide good support, especially in complex diagnoses like melanoma.