Anomaly detection in medical images using convolutional auto encoders: Blood smear images as a case study
Sujatha Kamepalli, Srinivasa Rao Bandaru, Kuruva Vamsi · 2025
Medical imaging is essential for diagnosing a wide range of medical conditions, enabling doctors to carefully examine and interpret images to determine patients’ health status. However, as the volume of medical images escalates, there&s;s a pressing need for automated systems adept at pinpointing anomalies within these images. This paper introduces an anomaly detection method employing Convolutional Auto Encoders (CAEs), tailored specifically for medical images. The methodology involves training a CAE on a dataset of normal blood smear images, utilizing it to reconstruct new images. An image failing accurate reconstruction is flagged as an anomaly. We employ two metrics, namely reconstruction error and Kernel density estimation (KDE) values, to discern anomalies. Evaluating our approach on a dataset inclusive of both normal and abnormal blood smear samples, we demonstrate its efficacy with an F1-score of 0.95. Our approach bears significant potential, including early disease detection, alleviating radiologists’ workload, and enhancing diagnostic accuracy. By serving as a diagnostic aid for pathologists, our model facilitates precise and efficient identification of anomalies in blood smear images.