Developing an AI project

Abstract

Artificial intelligence applications can very powerful in areas of speech recognition, image processing and identification, medical diagnosis and clustering to name a few. There is a perception that developing your own artificial intelligence (AI) application can be a daunting task, requiring in-depth knowledge and programming skills. This is not entirely true since many desktop and laptop systems have computing power that can accommodate machine-and deeplearning development, the available options for code development and a broad support base. A generic guide in developing a platform for AI project development is presented. R ESUM E Les applications d'intelligence artificielle peuventêtre tr es puissantes dans les domaines de la reconnaissance vocale, du traitement et de l'identification des images, du diagnostic et du regroupement, pour n'en nommer que quelques-uns. Il existe une perception selon laquelle le d eveloppement de votre propre application d'intelligence artificielle (apprentissage machine ou r eseau neuronal) peut constituer une tâche tr es difficile, exigeant des connaissances et des comp etences en informatique approfondies. Ce n'est pas enti erement vrai, puisque plusieurs ordinateurs de bureau et portables disposent d'une puissance de calcul permettant d'accommoder le d eveloppement de l'apprentissage machine et par r eseau neuronal, les options de codage et une vaste base de soutien. Nous pr esentons ici un guide g en erique pour l' etablissement d'une plateforme de d eveloppement de projets d'IA.

Discriminative deep learning based benignity/malignancy diagnosis of dermatologic ultrasound skin lesions with pretrained artificial intelligence architecture

“…DLA algorithms are usually pretrained with nonmedical images 20 . Large image databases such as Imagenet 12 permit initial weights preadjustment in order to make borders or other basic features of images more easily recognizable by algorithms 20 .…”

“…DLA algorithms are usually pretrained with nonmedical images 20 . Large image databases such as Imagenet 12 permit initial weights preadjustment in order to make borders or other basic features of images more easily recognizable by algorithms 20 . However, in the case of ultrasound, some authors suggest that initial weight adjustments should be made with the same kind of images to improve system accuracy 21 .…”

“…Another controversial aspect of training DLAs is the quantity and quality of images necessary. Most publications suggest the use of a large number of images in order to fine‐tune algorithms 20 . In our DLA, we used a limited number of images of seven categories and performance was similar to expert diagnosis.…”

“…The aphorism of “garbage in‐garbage out” is in no way improved by increasing the number of inputs without any sequential logic. In that sense, training of specific application DLAs should be done with the help of expert trainers in that application in order to make machine learning more systematic and to interpret failures of DLAs 20–21 …”

“…In that sense, training of specific application DLAs should be done with the help of expert trainers in that application in order to make machine learning more systematic and to interpret failures of DLAs. [20][21] The main limitations of our study are the limited number of images used, the restricted categories for DLA training (which exclude other diagnoses), and the possibility that other architectures not trained would be more efficient than EfficientNet B4. However, taking into account that these categories are the most frequently diagnosed in general dermatologic ultrasound, and that even with a limited number of images, efficiency for malignancy/benignity diagnosis was comparable to expert malignancy/benignity diagnosis.…”