Tuberculosis (TB), caused by Mycobacterium tuberculosis (Mtb), is a significant health concern, with early detection playing a crucial role in reducing mortality. The similarity between TB and lung cancer poses a challenge for radiologists in avoiding misdiagnosis during chest X-ray inspections. In recent years, applying deep learning techniques to medical image classification has shown promise due to the various nature of datasets. In developing nations, screening examinations and early diagnosis would benefit substantially from an affordable, effective automated procedure. Thus, we present a novel deep learning architecture to facilitate the detection of TB by integrating multisource data from deep learning-based methods. The novel framework for TB classification using Convolutional Autoencoder with EfficientNet (TB-CARE) involves the utilisation of a convolutional autoencoder for feature extraction, an affinity propagation (AP) clustering method for selecting templates, and an enhanced EfficientNet (EEffNet) for classification. Extensive tests are performed on five freely accessible datasets. The results of our methodology surpassed those of previous approaches, demonstrating its practicality for real-world applications. By leveraging deep learning models within the ensemble method, TB classification achieves a notable accuracy in testing as 99% for the Belurus dataset and 83% on Shenzhen, 99% on the TBX dataset and 60% to 70% for the Montgomery dataset and National Institutes of Health (NIH), where various classes with high imbalance affect the model's performance.