METHOD FOR AUTOMATED DETECTION OF PATHOLOGICAL CHANGES IN MACULAR REGION BASED ON ANALYSIS OF CROSS-SCANS OF OPTICAL COHERENCE TOMOGRAPHY USING ARTIFICIAL NEURAL NETWORKS Russian patent published in 2025 - IPC A61B3/12 

FIELD: medicine.

SUBSTANCE: invention relates to medical decision support systems in ophthalmology and can be used for automated detection and differentiation of pathological changes in the macular region based on the analysis of an array of cross-scans of optical coherence tomography (OCT) using artificial neural networks (ANN). Based on patient examination results, initial set of OCT cross-scans with size of 1645 × 641 pixels uniformly distributed over 11 classes, where the first class is represented by standard scans, and ten types of scans according to classes of the most common macular pathologies, on which visual signs of pathologies are marked. Then an array of files for validation of the ANN model is formed, which is 20% of the dataset, which is not involved in training. Cross-scan preprocessing is carried out, which consists in cutting from each image of the target informative fragment with size of 1440 × 600 pixels, and the obtained set of 11 classes of target fragments of colour images is used as a dataset for training the ANN model. Obtained informative fragments of the cross-scans are scanned vertically, changing the X coordinate from 1 to 1440, and the obtained files are processed, namely: obtained fragment is divided into component colours: Red, Green, Blue, and an array of integers corresponding to RGB colours of each pixel is formed. Obtained information is marked, which consists in selecting from the obtained records only those corresponding to any variant of the norm or pathology, indicating the class of pathology and the value of the coordinate X corresponding to the given pathology; information on RGB colours is averaged using a base, for which a value equal to 6 pixels is taken. Further, a dominant colour is selected for each base length unit, and the colours are encoded, and averaging is performed using the formula. Architecture of the ANN model is then formed and trained. When training the ANN model, an Adam optimizer is used, which uses the average value of the second moments of gradients, after validation of the ANN model, a table of alternations of dominant colours and their durations is obtained, which is information on the type of anomaly (or a combination of anomalies), its dimensions and coordinates, and obtaining normal images or an image with contours of macular pathology or a combination of pathologies, in particular location, area and number of signs. Method enables to optimize the process of making medical decisions in ophthalmology due to the possibility of detecting the presence of not only macular pathologies, but also combinations thereof, which helps to reduce the medical load due to accurate diagnosis and tracking changes in the state of the retina for selecting the optimal therapeutic approach to the patient.

EFFECT: use of a dynamic template enables to take into account all currently known macular anomalies and combinations thereof, the coordinates of which can be at different height of the image, and the absence of the need to use CNN, which enables to significantly reduce the amount of random access memory and make the invention more computationally efficient.

1 cl, 13 dwg, 1 tbl