Leukemia diagnosis requires highly skilled pathologists, as it is often based on the morphology of microscopic blood samples from the patient. In this context, the segmentation of microscopic blood cells plays a vital role in improving the final diagnosis of leukemia. This study applies an enhanced segmentation methodology, which includes optimizing multi-level Otsu for microscopic image segmentation followed by a range of machine learning and deep learning classifiers. The Acute Lymphoblastic Leukemia Database (ALL-IDB2) dataset is used for experimentation, both with and without segmentation, to demonstrate the effect of Learning-enthusiastic-based Teaching–Learning Optimization (LebTLBO) segmentation on leukemia classification. Initially, a Random Forest (RF) classifier is used, followed by a hybrid methodology employing VGG-16, Xception, and MobileNetV2 as feature extractors, with RF as the final classifier. Performance measures such as precision, recall, F1-score, accuracy, and the confusion matrix are considered. The accuracy of classification increases with LebTLBO segmentation compared to the raw dataset without segmentation. For the RF classifier, accuracy improves from 0.60 to 0.62 after segmentation. Similarly, accuracy improvements are observed, increasing from 0.68 to 0.73, 0.70 to 0.75, and 0.80 to 0.85 with the hybrid approaches of RF plus VGG-16, RF plus Xception, and RF plus MobileNetV2, respectively.