SOFTWARE ARCHITECTURE OF A NEURAL NETWORK IMAGE ANALYSIS SYSTEM FOR PRECLINICAL DETECTION OF AUTISM USING CLAUD TECHNOLOGIES

Abstract

The article considers the problem of preclinical screening risk detection of autism spectrum disorders and justifies the need for objectified, reproducible and scalable tools capable of supporting a specialist at the stage of primary screening. The software architecture of a neural network image analysis system using cloud technologies is proposed, which provides a managed data and model lifecycle, versioning, artifact storage, event logging and quality control. The architecture is structured into data and data management subsystems, a processing and training pipeline, as well as analysis services, access to which is provided through an API gateway with authentication and access control.

The proposed system is based on a neural network approach to preclinical risk detection of autism spectrum disorders, which implements three stages: additional training of a neural network model on the target dataset, photo classification and generation of explanations, including visual significance maps and semantic textual justification of the decision made. Experimental verification was performed on the ViT architecture model. The software example obtained a consistent probability distribution with an ASD risk score of about 0.724, as well as explanatory outputs in the form of key facial areas and morphometric indicators. The training results show that the best overall quality is achieved at the 5th epoch, where Val loss is 0.3506, Val acc 0.885, and MCC 0.771, while a further increase in the number of epochs leads to a decrease in the indicators.

The results obtained confirm the suitability of the proposed architecture for screening applications and determine the directions of further validation on independent, representative samples. The solution is focused on secure work with sensitive data due to centralized storage, auditing, and the ability to monitor changes in input data, which creates the basis for regulated updating of models in a cloud environment.