An Adaptive Query Approach for Extracting Medical Images for Disease Detection Applications

Different applications heavily benefit from automatic deep learning including image classification, segmentation, and analysis. It significantly adds value to clinical systems through computer-aided detection, curing planning, diagnosis, and therapy through the acquisition of the most informative images. However, this deep learning approach faces one of the main hurdles in image processing: the necessity of a large, labeled dataset. Actually, such requirements in medical image analysis applications are considered excessively costly to acquire. Active learning methods can mitigate such issues by reducing the number of annotated images while raising the model’s performance. This paper introduces an active learning framework based on a novel sampling technique, where it queries the unannotated samples that behave differently from current training set samples. The adaptive sampling method is optimized by stochastic gradient descent approximation. This optimization leads to the construction of an adaptable and robust system that meets the needs of medical control systems. Moreover, such novelties contribute to a respectful enhancement of the model’s deep network performance when training over a few numbers of annotated images to reach underlying accuracy. The proposed structure outperforms other AL methods, as proved by the experimental results using stochastic gradient descent optimization technique over Skin Cancer, Pediatric Pneumonia, and COVID-19 datasets, which achieved an accuracy of 72.5%, 90%, and 90.5 using only 42.8%, 8%, and 5% human-labeled training data, respectively.