Dermatopathology of Malignant Melanoma in the Era of Artificial Intelligence: A Single Institutional Experience

Abstract

The application of artificial intelligence (AI) algorithms in medicine could support diagnostic and prognostic analyses and decision making. In the field of dermatopathology, there have been various papers that have trained algorithms for the recognition of different types of skin lesions, such as basal cell carcinoma (BCC), seborrheic keratosis (SK) and dermal nevus. Furthermore, the difficulty in diagnosing particular melanocytic lesions, such as Spitz nevi and melanoma, considering the grade of interobserver variability among dermatopathologists, has led to an objective difficulty in training machine learning (ML) algorithms to a totally reliable, reportable and repeatable level. In this work we tried to train a fast random forest (FRF) algorithm, typically used for the classification of clusters of pixels in images, to highlight anomalous areas classified as melanoma “defects” following the Allen–Spitz criteria. The adopted image vision diagnostic protocol was structured in the following steps: image acquisition by selecting the best zoom level of the microscope; preliminary selection of an image with a good resolution; preliminary identification of macro-areas of defect in each preselected image; identification of a class of a defect in the selected macro-area; training of the supervised machine learning FRF algorithm by selecting the micro-defect in the macro-area; execution of the FRF algorithm to find an image vision performance indicator; and analysis of the output images by enhancing lesion defects. The precision achieved by the FRF algorithm proved to be appropriate with a discordance of 17% with respect to the dermatopathologist, allowing this type of supervised algorithm to be nominated as a help to the dermatopathologist in the challenging diagnosis of malignant melanoma.

Artificial Intelligence in Melanoma Dermatopathology: A Review of Literature

“…Cazzato et al 20 trained a fast random forest algorithm to classify clusters of pixels from digitized H&E slides as belonging to malignant melanoma or dysplastic nevi. The fast random forest algorithm was discordant with the dermatopathologist 17% of the time.…”

Artificial Intelligence Applied to a First Screening of Naevoid Melanoma: A New Use of Fast Random Forest Algorithm in Dermatopathology

“…The algorithm applied for feature extraction was the fast random forest (FRF) algorithm, a powerful supervised machine learning algorithm that optimizes the performance of the random forest (RF) algorithm with respect to computational speed and classification accuracy: it defines the best decision tree split condition step by step, thus avoiding unnecessary computations, extracting clusters of pixels appertaining to similar classes [1,14,15]. Specifically, the FRF algorithm implements Weka libraries and has been primarily used in industrial applications [15] and successively applied to medical images [13]. Theoretically, a similar class is identified by a group of pixels with different greyscale intensities and specific distances.…”

“…In this paper, after a previous work on MM [13], we present preliminary data on the training of an ML algorithm called fast random forest (FRF) which is applied to a dataset of histopathological images of NM; we discuss its strengths and limitations and provide a critical prospective for the near and distant future.…”

Skin Lesion Detection Using Hand-Crafted and DL-Based Features Fusion and LSTM

“…The above-mentioned experimental results indicate that our proposed skin lesion detector is a more effective system than existing techniques. In [ 53 ], the authors employed the fast random forest (FRF) algorithm to identify the area affected by malignant melanoma. They achieved 17% precision discordance with the pathologist’s results.…”