Thyroid Disease Classification using Hybrid CNN-BiLSTM Models: A Comparative Study with and Without Attention Mechanism

The tissue is the most frequently involved endocrine organ and therefore it requires an accurate and early detection for the treatment and management. Such diagnosis methods, while formally being a set of biochemical analyses and physical examinations, are subjective and rely on repetitive tests and visual scoring, necessitating advanced computer aided diagnostic schemes. This work presents a comparison of two hybrid deep models, CNN-BiLSTM and attention CNN-BiLSTM, in the context of thyroid disease classification. The CNN layer extracts spatial features from the input and the BiLSTM layer captures long-term dependency of patterns within thyroid function tests. For better feature selection and higher classification accuracy, the attention mechanism is introduced in the second model. To address the class imbalance in dataset SMOTE is used for over-sampling. Experimental results of the attention-based CNN-BiLSTM model are better than the baseline with more precision, recall and F1-score. Experiments demonstrate that the attention mechanism enhances the interpretability and classification performance of the model. This paper demonstrates the potential of attention-augmented deep learning models for reliable automatic thyroid disease diagnosis systems to improve clinical decision-making and patient handling.