Cervical cancer is still one of the major causes of cancer death in women around the world. Accurate and early classification of precancerous lesions from images of cervical cells is crucial for the enhancement of patient outcomes. Based on MMT, this paper introduces a new hybrid model named TabNet-PANDCG for improving the classification performance, severity-aware ranking, and interpretability of models. This paper presents TabNet-PANDCG, a novel hybrid model built on top of MMT to boost classification performance, severity-aware ranking, and model interpretability. The proposed method extracts the morphological features, intensity features and texture features from cervical cell images to generate tabular structured images, in contrast with the conventional method, which processes raw images directly. The image-derived features are then passed into Multi-Task TabNet with a sparse attention mechanism that performs the task of dynamically selecting the relevant features and uses it to simultaneously classify the lesion and predict the lesion's severity. The proposed PANDCG metric is a ranking evaluation that assigns greater weight to the clinically relevant, high-severity lesions that rank highly, giving a more meaningful evaluation that fits more with the clinical focus. The framework was tested using a publicly available dataset containing 917 Pap images from single-cell images, across seven diagnostic categories, namely Herlev. Experimental results show that TabNet-PANDCG has the greatest performance with 97.39% accuracy, 93.66% macro precision, 92.17% macro recall and 92.14% macro F1 score, outperforming some of the state-of-the-art techniques. Additionally, the incorporation of PANDCG facilitates clinically relevant ranking assessments, and TabNet's attention-based feature selection further promotes transparency and trust in the model. The framework exhibits great promise for computer-aided diagnosis systems in cervical cancer screening.