Abstract:

In this work we tried to train an AI algorithm using basic histopathological criteria that indicate and differentiate with a good probability a malignant melanoma from a severe dysplastic nevus (atypical nevus), and we provide information on the results obtained starting from routine histopathological images. The artificial intelligence image processing algorithm used to classify and to enhance anomalies contained in the microscope image is the Fast Random Forest (FRF). The learning process of the algorithm is based on a preliminary classification of cluster of pixels of the same image including possible Melanoma’s areas: the preliminary identification of Melanoma morphological features, represents the labelling approach typical of machine learning supervised algorithms. The FRF testing provides as output the processed image with colored enhanced Melanoma pixel clusters (each class selected in the learning step is represented by a color), probabilistic maps (high probability highlighted by white to identify an anomaly in a specified image region), and algorithm performance indicators (precision, recall, and Receiver Operating Characteristic -ROC- curves [5]). The optimized hyperparameters and filter properties applied for the image FRF processing (features training) are [5]: Gaussian blur filter, Hessian matrix filter, membrane projections, membrane thickness equals to 1, membrane patch size equals to 19, minimum sigma equals to 1, maximum sigma equals to 16. For five pixel clusters of the same dimensions, occurs a number of about 300 instances (computational cycles) to achieve the maximum precision (equals to 1), with a computational cost of about 2 minutes using a processor Intel(R) Core(TM) i5-7200U CPU, 2.71 GHz. The minimum recall performance parameter (near to 0) is achieved about 392 instances. The ROC curve (representing in the plane the true positive rate versus the false positive rate) is matching with the ideal curve of a perfect classifier.

Financial Disclosure:
No current or relevant financial relationships exist.

Menu

Angelo Vacca