Outliers and anomalies in training and testing datasets for AI-powered morphometry -Evidence from CT scans of the spleen

Yuriy Vasilev^1,2Anastasia Pamova^1*Tatyana Bobrovskaya^1*Anton Vladzimirskyy^1,3Olga Omelyanskaya¹Elena Astapenko¹Artem Kruchinkin¹Kirill Arzamasov^1,4

• ¹Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Health Care Department, Moscow, Russia
• ²National Medical and Surgical Center named after N.I. Pirogov of the Ministry of Health of the Russian Federation, Moscow, Russia
• ³I.M. Sechenov First Moscow State Medical University of the Ministry of Health of the Russian Federation (Sechenov University), Moscow, Russia
• ⁴Moscow Technical University - MIREA, Moscow, Russia

Introduction: Creating training and testing datasets for machine learning algorithms to measure linear dimensions of organs is a tedious task. There are no universally accepted methods for evaluating outliers or anomalies in such datasets. This can cause errors in machine learning and compromise the quality of end products. The goal of this study is to identify optimal methods for detecting organ anomalies and outliers in medical datasets designed to train and test neural networks in morphometrics.Methods: A dataset was created containing linear measurements of the spleen obtained from CT scans. Labelling was performed by three radiologists. The total number of studies included in the sample was N = 197 patients. Using visual methods (1,5 interquartile range; heat map; boxplot; histogram; scatter plot), machine learning algorithms (Isolation forest; Density-Based Spatial Clustering of Applications with Noise; K-nearest neighbors algorithm; Local outlier factor; One-class support vector machines; EllipticEnvelope; Autoencoders), and mathematical statistics (z-score, Grubb's test; Rosner's test). Results: We identified measurement errors, input errors, abnormal size values and non-standard shapes of the organ (sickle-shaped, round, triangular, additional lobules). The most effective methods included visual techniques (including boxplots and histograms) and machine learning algorithms such is OSVM, KNN and autoencoders. A total of 32 outlier anomalies were found.Discussion: Curation of complex morphometric datasets must involve thorough mathematical and clinical analyses. Relying solely on mathematical statistics or machine learning methods appears inadequate.