Attention-enhanced deep learning for cervical cytology: combining convolutional networks with multi-head attention and fuzzy logic.

Purpose: Cervical cancer continues to be one of the leading causes of death among females worldwide, and thus early diagnosis by using more advanced diagnostic procedures is crucial. The conventional Pap-smear procedure is accurate but subject to human error; thus, computerised, standardised, and automated diagnosis becomes imperative. Herein we present a novel framework of a fuzzy distance-based ensemble of convolutional neural networks (CNNs) for efficient cervical cancer classification from Pap-smear images. Material and methods: The proposed approach integrates 5 models of CNN - Simple CNN, InceptionV3, Xception, Xception with Attention, and Inception Attention - via attention mechanisms to advance feature learning. A fuzzy distance-based aggregator function is introduced to fuse the predictions of these models optimally as per Euclidean, Manhattan, and cosine distance measures. Four advanced pre-processing techniques - wavelet denoising, contrast-limited adaptive histogram equalisation (CLAHE), background correction, and Laplacian sharpening - are employed to construct a cleaner dataset with enhanced image sharpness and segmentation. Results: Experimental outcomes prove that the model is significantly better than state-of-the-art approaches, with an accuracy of 94% on the original dataset and 98.3% on the pre-processed dataset. Conclusions: The method suggested herein has better noise robustness, interpretability through fuzzy logic, and automatic adaptation to various CNN frameworks without fine-tuning. These results acknowledge the promise of fuzzy logic-based CNN ensembles to improve machine-based cervical cancer diagnosis, which could be mapped to better and scalable diagnostic instruments in medical imaging.