The Anti-Nuclear Antibodies (ANA) test, utilizing Human Epithelial type 2 (HEp-2) cells in the Indirect Immuno-Fluorescence (IIF) assay, is widely regarded as the gold standard for detecting Connective Tissue Diseases. However, the manual analysis of HEp-2 images remains a time-consuming and subjective process, creating a growing need for reliable, computer-assisted solutions. In this study, we propose an innovative platform leveraging transfer learning with pre-trained deep learning models to address key challenges in HEp-2 image analysis. Our approach combines: Unsupervised deep feature extraction to describe HEp-2 images effectively. A novel feature selection technique tailored for unbalanced datasets. Independent testing on datasets from two distinct hospitals, addressing cross-hardware compatibility challenges. To improve the trustworthiness and interpretability of our platform, we introduce: A modified Gradient-Weighted Class Activation Mapping (Grad-CAM) method for regional explainability, enabling deeper insights into classification decisions. A new sample quality index based on the Jensen-Shannon divergence, which enhances reliability by quantifying sample heterogeneity. Our results demonstrate state-of-the-art performance in both intensity and ANA pattern recognition, surpassing existing methods. Notably, our platform eliminates the need for traditional cell segmentation, relying instead on statistical sample analysis. This makes our method more robust, versatile, and broadly applicable across varying clinical settings. Looking ahead, we aim to expand our platform to address the challenge of mitotic spindle recognition, enabling the analysis of mixed ANA patterns and further advancing its potential in clinical diagnostics.