Architecture of our method. In the ITS, each labeled sample undergoes convolutional neural network processing to train an initial model. Subsequently, domain experts annotate the K most challenging samples, as determined by the initial model’s judgments. During the ALS, we employ the combined set of (M + K) labeled samples to train an active training model. From this model, we select the top V samples with high-confidence predictions and assign pseudo-labels to them. In the SSLS, we utilize the expanded dataset of (M + K + V) samples to train a semi-supervised training model. This process is repeated for a total of R iterations to obtain the final results.