A Precise Cervical Cancer Classification in the Early Stage Using Transfer Learning-Based Ensemble Method: A Deep Learning Approach

Cervical cancer stands out as one of the deadliest forms of cancer affecting women, underscoring the significance of timely detection for effective treatment, a principle applicable to all cancer variants. Although the Pap smear test stands as the benchmark for this type of cancer diagnosis, the accuracy of this diagnosis depends on the skill and attentiveness of the healthcare provider. Considerable efforts have been directed toward leveraging artificial intelligence-based computer-aided diagnostic techniques to augment traditional cervical cancer detection processes. Numerous deep learning models have been developed by scholars using these approaches to classify cervical cancer. However, these models necessitate preliminary segmentation steps from Pap smear slides for isolating cervical cells. This undertaking is not only intricate but also resource-intensive and time-consuming. Furthermore, the challenges of cervical cancer classification using deep learning also encompass issues related to the adequacy and caliber of data, alongside the disparities in the dimensions, contours, and visual attributes of cervical cancer images. In response to these obstacles, we present an innovative ensemble architecture founded on cutting-edge transfer learning. This approach involves the fusion of two of the most recent and potent convolutional neural networks, namely InceptionV3 and EfficientNetV2S. Our proposed ensemble model incorporates supplementary layers, including concatenate, flatten, batch normalization, multiple dense layers, PReLU activation, and a concluding softmax classifier layer. The objective is to achieve optimal accuracy on the SIPaKMeD Pap smear dataset, which is benchmarked, without the necessity for any segmentation techniques. Furthermore, we have integrated the Adam optimizer into our model to mitigate issues of both overfitting and underfitting. In terms of detecting cervical cancer, our model surpasses all the contemporary models developed by researchers using the SIPaKMeD dataset in accuracy. Our model demonstrates a remarkable accuracy of 0.9998, precision of 0.9998, and recall of 0.9998 and achieves an AUC score of 1.000.